SUNfowerNext

Towards a composite road

safety performance index
Fred Wegman (SWOV)
Jacques Commandeur (SWOV)
Etti Doveh (Technion)
Vojtech Eksler (CDV)
Victoria Gitelman (Technion)
Shalom Hakkert (Technion)
David Lynam (TRL)
and Siem Oppe (SWOV)
S
U
N
f
o
w
e
r
N
e
x
t
:

T
o
w
a
r
d
s

a

c
o
m
p
o
s
i
t
e

r
o
a
d

s
a
f
e
t
y

p
e
r
f
o
r
m
a
n
c
e

i
n
d
e
x
I
S
B
N
:

9
7
8
-
9
0
-
7
3
9
4
6
-
0
5
-
7






SUNflowerNext:
Towards a composite road safety performance index



Fred Wegman (SWOV), Jacques Commandeur (SWOV), Etti Doveh (Technion),
Vojtech Eksler (CDV), Victoria Gitelman (Technion), Shalom Hakkert (Technion),
David Lynam (TRL), Siem Oppe (SWOV)















SWOV Institute for Road Safety Research, the Netherlands









Project co-funded by the European Commission within the Sixth
Framework Programme (2002 -2006)








Project co-financed by the European Commission,
Directorate-General Transport and Energy
Report documentation
Title: SUNflowerNext: Towards a composite road safety
performance index

Authors: Fred Wegman (SWOV), Jacques Commandeur (SWOV), Etti
Doveh (Technion), Vojtech Eksler (CDV), Victoria Gitelman
(Technion), Shalom Hakkert (Technion), David Lynam (TRL),
Siem Oppe (SWOV)

Keywords: Safety, policy, methodology, indicator, benchmark, statistics,
traffic, injury, fatality, accident rate, Europe.

Number of pages: X + 126 + 48
Price: 30,-

Published by: SWOV, Leidschendam, 2008


ISBN: 978-90-73946-05-7
NUR: 976
URL: http://sunflower.swov.nl
http://www.swov.nl





















SWOV Institute for Road Safety Research
P.O. Box 1090
2260 BB Leidschendam
The Netherlands
Telephone: +31 70 317 33 33
Telefax: +31 70 320 12 61
E-mail: info@swov.nl
Internet: www.swov.nl
Foreword
You can't manage what you can't measure
Robert Kaplan, Harvard Business School

Measuring the way to knowledge ("Door meten tot weten")
Heike Kamerlingh Onnes, Dutch Nobel Prize winner, 1913

In 2002, the first SUNflower report was published. This report compared the road
safety in three countries: Sweden, the United Kingdom, and the Netherlands. The
comparison was made as an attempt to identify the similarities and differences
between these countries, not just with respect to the numbers of crashes and
victims. Particularly the factors, circumstances, developments that have an influence
on the risk of a crash and the severity of its outcome were investigated. The
underlying thought was that the findings could be helpful in the possibility to learn
from each other. Although the mortality and the risks are approximately equal in the
three countries, large differences were found in the ways road safety improvements
were tackled. Furthermore, although the mortality was approximately equally high
(or, rather, equally low), various large differences were found between the
explanations: in the United Kingdom, for example, the fatality risks for motorcyclists
and pedestrians showed to be high as opposed to that in Sweden and the
Netherlands, as it did for car occupants in Sweden compared to the other two
countries, and for moped riders in the Netherlands. The long list of recommen-
dations showed that with the analysis method used it was indeed possible for
countries to learn from each other.

Understandably, this result was reason to enlarge the SUNflower range and to
attempt a further deepening: the SUNflower+6 study was initiated and reported on in
2005. Nine European countries participated in this study: the original three
countries, three countries in Southern Europe (Greece, Portugal and Spain, with a
special position for Catalonia), and three Central European countries (Hungary,
Slovenia and the Czech Republic). This study also resulted in very interesting
insights and useful recommendations. However, this study also showed that the
more the countries differ between themselves, the harder it becomes to interpret the
comparisons. A tendency arose to make three comparisons of three countries and
not so much compare all nine countries.

The study also resulted in a first design of a road safety footprint. A footprint was
defined as a representation of the road safety status of a country. A footprint
contains a combination of indicators, measured as a snapshot in time or a time
series. This footprint was found to be an interesting concept deserving further
elaboration. The present SUNflowerNext study has carried out this task.

This elaboration was done as part of the SafetyNet project. This project aims to build
the framework of a European Road Safety Observatory, which will be the primary
focus for road safety data and knowledge, as specified in the Road Safety Action
Programme (EC, 2003).

SUNflowerNext is aimed at the development of a knowledge-based framework for
comprehensive benchmarking of road safety performances and developments for a
country or other sub-national jurisdictions. An explorative method was used to
accomplish this. In this project we limited ourselves to the use of readily available
data; no additional data was collected.
I

The SUNflower approach and its proven benefits (data driven, comprehensiveness
through the road safety pyramid, science-based understanding of differences
between benchmark values, identification of potential improvements of perform-
ances) aims at presenting the best relevant knowledge and available data, and
introduces best practices for benchmarking of countries or sub-national jurisdictions.

Over time, this study has explored different directions. They address different
aspects of ranking or benchmarking of safety performances of countries and
regions/cities in countries. The explorative character of the work and the immaturity
of the different developments so far make a conclusion about obvious dissimilarities
between the different chapters unavoidable. If this approach is to continue, and we
will certainly support this decision, we will certainly pay more attention to this issue.

Three groups worked on this report. One group consists of Shalom Hakkert, Victoria
Gitelman and Etti Doveh of Technion in Haifa, Israel. This group mainly investigated
the possibilities for the development of a composite index in Chapter 3. A second
group focused at the sub-national level and consisted of David Lynam (UK) and
Vojtech Eksler (Czech Republic). Their contribution can be found in Chapter 6.
Finally, the Dutch group, formed by Jacques Commandeur, Siem Oppe and Fred
Wegman worked on the remaining chapters. But I would wrong everyone involved
by only linking the researchers to individual chapters. This explorative study's quest
for ways that enable a good comparison between countries did not follow an easy
path. No appropriate examples were available, although similar efforts were
undertaken in other social disciplines (see also Chapter 2). However, the specific
character of the road safety issue required the exploration of many new paths.
Together we explored this terra incognita. And I can tell you a secret: that path was
not always straight, but was often bendy; it sometimes was hard to even find a path
at all; we were not always a tight-knit group and did not always immediately agree at
each fork of the path which branch would be the shortest and fastest one to our
destination. But together we reached the finishing line. Our process brought a quote
to our minds: "If you want to go quickly, go alone. If you want to go far, go together".
Al Gore quoted here an African proverb in his Nobel Lecture in Oslo, 2007.

This is my third foreword to a SUNflower report. Each time the same emotions rise
up when I see a group of eminent researchers driven to excellent performance on
the basis of their thorough knowledge of road safety and their research in this field.
This is the place to express my gratitude to the entire team. My gratitude also goes
to two SWOV employees whose efforts ensured that an excellent version of the
report could be published in print. As on many other occasions, Marijke Tros and
Hansje Weijer have significantly contributed to the quality of the final product. I also
extend my gratitude to the SafetyNet consortium for welcoming us within their ranks,
and to the 'reviewers' who made a useful contribution to utmost quality.

I very much hope that the set-up for benchmarking the safety performance of
countries (or sub-national jurisdictions) and the idea of defining a composite road
safety performance index (SUNflower index) will be realized, will be measured and
published annually, and will hence provide a firm basis for road safety improvements
in the EU Member States and a possibility to continue to learn from one another.

Fred Wegman
II
Content
Foreword .......................................................................................................... I
Content ........................................................................................................ III
Executive summary................................................................................................VII
1. Integration of SafetyNet and SUNflower..................................................... 1
1.1. Introduction.....................................................................................................1
1.2. Aim of the study..............................................................................................3
1.3. The SUNflower approach ...............................................................................3
1.3.1. System boundaries and external influences.....................................5
1.3.2. Road safety management: safety measures and programmes........6
1.3.3. Structure and culture........................................................................7
1.3.4. Vertical relationships between the different layers .........................10
1.3.5. Social costs ....................................................................................11
1.4. Structure of the report...................................................................................12
2. Benchmarking by using road safety indicators....................................... 14
2.1. Benchmarking road safety performances.....................................................14
2.2. Performance indicators for road safety.........................................................15
2.3. Road safety: towards a composite performance index.................................18
2.4. Performance indicators for road safety policies and their implementation ...20
2.4.1. Benchmarking policy performance .................................................21
2.4.2. Benchmarking implementation performances ................................22
2.5. Conclusions..................................................................................................25
3. Designing a composite index for road safety.......................................... 27
3.1. Basic indicators ............................................................................................29
3.2. Data collection..............................................................................................32
3.3. Method of analysis........................................................................................32
3.4. Summary of findings.....................................................................................35
3.5. Discussion and conclusions .........................................................................40
3.5.1. Comparisons of countries' rankings ...............................................41
3.5.2. Identification of groups of countries................................................44
3.5.3. Behaviour of basic indicators .........................................................46
3.5.4. Towards the SUNflower road safety performance index................46
4. Grouping countries .................................................................................... 48
4.1. Grouping by safety experts based on the Sunflower+6 study ......................49
4.2. Singular Value Decomposition of traffic safety developments......................50
4.2.1. General description ........................................................................50
4.2.2. Fatality risk developments (fatalities per motor vehicle kilometre) in
13 European countries ...................................................................51
4.2.3. Fatality rate developments (fatalities per number of motor vehicles)
in 20 European countries ...............................................................54
4.3. Multiple Correspondence Analysis ...............................................................57
4.4. Grouping countries, based on three strategies.............................................61
III
5. .......................................................................Time series analysis
....................................................................................................... 63
5.1. Fatality and mortality developments in 20 European countries ....................64
5.2. SVD analysis of the fatality risk developments in 11 European countries....65
5.3. Trends in fatality risks and rates for individual countries ..............................67
5.3.1. Model structure...............................................................................67
5.3.2. Results for three countries .............................................................67
5.4. Trends of fatality rates for grouped countries ...............................................70
5.5. Analysis of disaggregate data for groups of countries..................................74
5.6. Conclusions and recommendations .............................................................82
6. Application of SUNflower at regional or city level................................... 83
6.1. Differences in application at the sub-national level.......................................84
6.1.1. Potential uses of sub-national comparison.....................................85
6.1.2. Understanding factors affecting sub-national comparison..............85
6.1.3. Applying the pyramid approach......................................................86
6.2. Role of (physical) structural differences........................................................90
6.3. Role of cultural factors..................................................................................95
6.4. Factors affecting measures and programmes ..............................................96
6.4.1. The SUNflower approach...............................................................97
6.4.2. Scope for extended SUNflower analyses.......................................97
6.5. Regional comparisons ..................................................................................98
6.5.1. Current practices and applications .................................................98
6.5.2. Avenues for further work ..............................................................101
6.6. City comparisons ........................................................................................103
6.6.1. Current practices and analyses....................................................103
6.6.2. Avenues for further work ..............................................................109
7. Conclusions and recommendations....................................................... 112
7.1. Benchmarking of road safety performances...............................................112
7.2. Indicators for road safety ............................................................................113
7.3. Towards a composite road safety performance index................................114
7.4. Time series analyses..................................................................................116
7.5. SUNflower at regional or city level?............................................................116
References ..................................................................................................... 119
Appendix 1. Final data set with initial and imputed values of basic
indicators .................................................................................... 127
Appendix 2. Detailed results of the five analyses......................................... 130
Appendix 3. Tools produced by Principle Component analyses for the
estimation of country scores .................................................... 151
Appendix 4. Tools produced by common factor analyses for the estimation
of country scores ....................................................................... 155
Appendix 5. Singular Value Decomposition (SVD)....................................... 159
Appendix 6. The latent risk model ................................................................. 161
Appendix 7. Fatality trends for individual countries .................................... 164
IV
Appendix 8. Disaggregate developments for groups of countries............. 167
Appendix 9. IAL - Local Accident Indicator................................................... 172
Appendix 10. Literature overview of studies on regional risk analysis........ 173
V

Executive summary
Background and aim
One of the aims of international cooperation in the field of road safety is to make
oneself familiar with performances and progress in other countries and to
understand if and how these can be of guidance to policymaking, in an adapted form
if appropriate. Comparisons can be a starting point to learn from each other.

The learning includes subjects such as monitoring and explaining road safety
developments, and gaining good insights in the impacts of interventions as a basis
for speeding up road safety improvements in one's country or jurisdiction.

Benchmarking is a process in which countries or sub-national jurisdictions evaluate
various aspects of their performance in relation to that of other counties or
jurisdictions, including the so-called 'best-in-class'. The benchmark results provide
countries or sub-national jurisdictions with information about others that can be used
as a basis for developing measures and programmes to increase their own
performance.

Two important tasks can be identified in this process:
1. defining the key components of a road safety performance and investigating if
and how these key components can be combined in a composite index;
2. finding a meaningful reference (best-in-class) and defining procedures for
identifying such a meaningful reference.

Comparing performances and, one step further, benchmarking performances seems
to be an appropriate approach for road safety. This approach should help us to go
beyond the rather traditional methods of comparing performances by only using
mortality rates or fatality rates or risks. Ranking countries by using only these rates
is a useful first step, but not very meaningful as a start to learn from each other.

The SafetyNet project aims to build the framework of a European Road Safety
Observatory, which will be the primary focus for road safety data and knowledge, as
was specified in the Road Safety Action Programme 2003. In the SafetyNet project it
was decided to develop a method of benchmarking road safety by using road safety
indicators. To this end, the SUNflower approach was used, more precisely the
information captured in the SUNflower pyramid and earlier attempts to elaborate on
this in developing the SUNflower footprint, as well as other SUNflower studies. We
gave this project the name SUNflowerNext.

Hence, the aim of the SUNflowerNext project is to develop a knowledge-based
framework for comprehensive benchmarking of road safety performances and
developments of a country or of sub-national jurisdictions.

SUNflowerNext has made use of existing data that was relatively easily available.
This ensured that the study could be carried out in a relatively short time. However,
one important concession needed to be made. Because this study used an
innovative approach with only existing data that was not always available, it was
decided to set up the research in such a way that all the steps required for
benchmarking a country's performance are taken, but to refrain from presenting the
actual results of the benchmark as they are of insufficient quality. The experiences
VII
gained from this study are such that SUNflowerNext's ambition benchmarking the
safety performance of countries - is realistic once reliable data is available.
Therefore, it is recommended to carry out this benchmarking in Europe in the near
future, to widely disseminate the results, and to consequently use them for policy
making in the European Member States.
Benchmarking of road safety performances
Benchmarking is a process in which actors evaluate various aspects of their
performance in relation to others, and to the so-called 'best in class'. In the
SUNflowerNext study we researched whether countries in the European Union
could all be placed in one class, or whether we should consider working with two or
more classes. Three procedures were used to find out whether meaningful groups
could be made: safety experts were asked to group countries, secondly, countries
were grouped based on road safety outcome indicators (grouping obtained with a
Singular Value Decomposition (SVD) of the annual fatality risks in the years 1980-
2003 of countries), and, thirdly, countries were grouped using general statistical data
from a Multiple Correspondence Analysis (MCA) about a country in the most recent
years.

In the SUNflowerNext project we concluded that it is better not to make comparisons
between all European countries as one group, but to attempt grouping comparable
countries and to then compare the countries within a specific group or class. The
results of the three methods have many points of agreement. The grouping results
have a preliminary character and it is recommended to elaborate on this topic before
coming to a final decision on the grouping. The approach explored in
SUNflowerNext could be used for this purpose.
Towards a composite road safety performance index
SUNflowerNext decided to develop an integral and comprehensive set of indicators
to measure the road safety performance of a country while including all information
in the SUNflower pyramid. SUNflowerNext distinguishes three types of indicator: the
road safety performance indicator, the implementation performance indicator, and
the policy performance indicator.

The first type of indicator captures a country's road safety quality. It has been
named the Road safety performance indicator. Other names such as outcome
indicator and product indicator are also used. In SUNflower the three top layers of
the SUNflower-pyramid are included: final outcomes (numbers of killed and injured),
intermediate outcomes (such as the safety performance indicator), and social costs.

The second type of indicator specifies the quality of the implementation of road
safety policies: the Implementation performance indicator. For this implemen-
tation quality indicator the term process indicator can also be used. Basically, this
indicator follows a vertical line in the pyramid linking 'safety measures and pro-
grammes', safety performance indicators and numbers of killed and injured people.

The third type of indicator deals with the quality of policy to improve road safety: the
Policy performance indicator. Here SUNflowerNext distinguishes two compo-
nents: the quality of conditions (strategies, programmes, resources, coordination,
VIII
institutional settings, etc.) and the quality of action plans and individual
(counter)measures) in the perspective of the ambitions expressed in road safety
targets.

There are several reasons why it is attractive to combine all information in one
indicator, a so-called composite index. A composite index includes all components
of the SUNflower pyramid, more specifically the three types of indicator. The pros
and cons of working with composite indices are rather well known and are presented
in the report. Three words can summarize the main characteristics: 'simplification,
quantification and communication'. Road safety will not be the first policy field to
successfully attempt to capture performance in one single value. To mention a few:
the Human Development Index, the Environmental Sustainability Index, and the
Overall Health System Index. Based on these examples it was decided to also
explore the opportunities for a composite index for road safety performance.

Weights based on statistical models were used to combine the basic indicators into
a composite index. Both Principal Component Analysis (PCA) and Common Factor
Analysis (FA) weighting were examined. Both methods group collinear indices to
form a composite index that captures as much as possible of the information that is
common among sub-indicators. The analysis was made on the data collected for 27
European countries. The composite index enables us to rank the countries in
accordance with their safety performance.

The analysis revealed that the countries' ranking based on the combination of
indicators is not necessarily similar to the traditional ranking of countries based only
on mortality rates or fatality rates. We believe that adding information on policy
performance and implementation performance to the ranking and grouping process
improves the results beyond the established methods and makes them more
comprehensible. Furthermore, it was observed that the indicators belonging to the
final outcomes and intermediate outcomes, both part of the road safety performance
indicator, are not uniform in their behaviour. The indicators that were found to be
more consistent and termed 'core set of basic indicators' are recommended for
future uses.

The general conclusion is that the design of a composite road safety performance
index, for example the SUNflower index in which relevant information from the
different components of the road safety pyramid has been captured and weighted, is
realistic and meaningful. In addition, such an index gives a more enriched picture of
road safety than a ranking only based on data on mortality or fatality rates, which is
common practice at present. Grouping countries using this process is promising and
seems to be preferable to simply ranking countries. Before defining the SUNflower
index and actually applying the results to policy making, two improvements should
be made: indicators must be developed for the Implementation performance
indicator and procedures must be developed to make available high quality and
comparable data for EU Member States.
Time series analysis
Safety developments are interesting because they may give us a better insight in
underlying forces and, hopefully, also in the effectiveness of road safety
interventions. Different approaches were used in this part of the study, among which
state space modelling. The first attempt to compare developments in fatality rates
IX
(fatalities per 10,000 motorized vehicles) and mortality rates (fatalities per 100,000
inhabitants) was made at a macroscopic level. Although European countries do
have a remarkably different history when it comes to the development of fatality rate
vs. mortality rate, our data suggests that all countries seem to be moving to the
same road safety position, although not at the same pace. Leading countries in the
field of road safety generally keep ahead of the other countries, albeit with
decreasing advantage.

Three types of disaggregate developments were compared (age, transport mode
and road type). In this comparison countries were grouped. Looking at the results of
the analyses, we may conclude that, although all European countries tend towards
the same aggregated or macroscopic level of road safety, there are important
differences between the individual countries as well as between groups of similar
countries. These differences relate to how they reach this level of road safety when
considering their focus on avoiding special types of accidents. In other words, the
general policies of improving road safety in different countries ultimately seem to
move towards the same safety level, but for different countries that level of road
safety is achieved at a different pace and in different ways.
Sub-national comparisons
There are two basic reasons for comparing the safety performance of sub-national
jurisdictions. In the first place, a ranking of relative performance of each area will be
very useful for comparison within countries. In the second place, it will provide better
understanding of the factors affecting safety improvement, so that safety
practitioners can achieve more effective programmes. This requires greater focus on
understanding how the effects of programmes are modified by the nature of the
safety problems faced by each area. Lessons can not only be learned from
comparison of areas within countries, but also from comparison of similar areas in
different countries.

The study clearly identifies factors which have effects on risks at a regional and local
level. Based on a literature review it was concluded that structural and cultural
differences, the bottom layer of the pyramid, can considerably affect road safety at a
regional and local level. The results of this part of the study are considered
sufficiently interesting for recommending continuation of this work in an international/
European project. In addition, it is recommended to use different approaches for
studies at both the regional and the urban level.

X
1. Integration of SafetyNet and SUNflower
1.1. Introduction
We can observe a growing interest and a growing number of activities in the field of
road safety internationally, more specifically in Europe. Activities supported by the
European Commission like, for example, including road safety research in the
European framework programmes, are instrumental in this process. European road
safety policies, as expressed for example in the EC's White Paper on European
transport policy (EC, 2001) and in the European Road Safety Action Programme
(EC, 2003) also encourage international cooperation. One of the more visible
activities is the recent establishment of a European Road Safety Observatory
(ERSO). This Observatory has different aims, among which monitoring progress
towards road safety targets and identifying best practices.

The SafetyNet project was initiated with the aim to build up the ERSO, paying
attention to three different areas: collecting and analysing data at a macroscopic
level (CARE, risk exposure data, and safety performance indicators), in-depth-data
(independent accident investigation and in-depth accident causation data) and
knowledge on road safety topics (www.erso.eu).

During the course of the SafetyNet project it was decided to incorporate the
SUNflower approach in SafetyNet in order to integrate different components of the
SafetyNet activities. This report reflects the work that was done to accomplish this
task.

The first SUNflower report (Koornstra et al., 2002), comparing road safety in
Sweden, the United Kingdom and the Netherlands, formulates the basic idea behind
the SUNflower approach: "A better insight into the development of policies and
programmes in these countries might conceivably identify key factors, which could
further improve current safety practice in each of them". From analysis and
diagnostic point of view, the SUNflower approach aims to identify strong and weak
points in the road safety performance of different European countries. The aim of
this approach is to determine underlying elements in the current policies and
programmes in EU Member States, to learn which of these elements make them
particularly effective in coping with the traffic safety problem, and thereby identify
policy improvements most likely to result in further casualty reductions. Of course, it
is also possible to use the opposite approach. From an intervention point of view
(how can road safety effectively be improved) the SUPREME project, for example,
identified and published best practices in road safety in the EU Member States (KfV,
2007).

Comparing three countries was found to shed very interesting light on the
performances of the countries. In many senses the countries differ a lot, but they are
the three countries with the highest road safety level in the world. We found that
these similar levels of safety were achieved through continuing planned
improvements over recent decades, that the targeted policy areas had been similar,
but implemented policies differed at a detailed level. In the second study, called
SUNflower+6, the number of countries was increased to nine (Wegman et al.,
2005). The positive outcome from the initial SUNflower study was more or less
repeated in the SUNflower+6 study, although the final conclusions were not as easy
to interpret as those in the SUNflower study. But comparing performances and
1
safety developments in three groups of countries with similar road traffic
backgrounds resulted in interesting and meaningful recommendations for all nine
countries.

The comparisons of nine countries made clear that just the comparison of countries
did not sufficiently generate the interest of the researchers and policymakers. It was
evident that learning from each other, and especially from the best performing
countries added an extra dimension to this approach. For that reason we introduced
the concept of benchmarking. Benchmarking is an action aimed to improve your
performance by learning from others through 1) identifying and 2) understanding,
and by 3) adapting outstanding practices from the countries which are considered to
be 'best-in-class'. This concept originates from business/the private sector, but can
also be applied, for example in comparing road safety performances between
countries.

Comparing or benchmarking countries in the field of road safety presupposes a set
of indicators, which together paint the whole picture. This set of indicators is called a
benchmark. A frequently used word as indicator needs some clarification to reach a
common understanding. In general terms, an indicator is a quantitative or a
qualitative measure derived form a series of observed facts that can reveal relative
positions (e.g. of a country) in a given area (Nardo et al., 2005). According to these
authors a composite indicator is formed when individual indicators are compiled into
a single index on the basis of an underlying model. The composite indicator should
ideally measure multi-dimensional concepts which cannot be captured by just a
single indicator, e.g. competitiveness, industrialization, sustainability, single market
integration, knowledge-based society, etc.

The road safety target hierarchy in the SUNflower approach (see also Figure 1.1)
was introduced to compare road safety performances of countries. This hierarchy
acknowledges the different aspects of road safety and road safety interventions.
These aspects can be measured when they are properly defined and can be
correctly measured by using different indicators. To make matters even more
complicated, the indicators at all levels of this hierarchy are also multi-dimensional.
To illustrate this point, final outcome indicators are expressed in terms of 'Number of
killed and injured'. However, Elvik (2008) suggests that this number (magnitude) is
only one of the nine possible characteristics. The other eight that are identified by
Elvik are: severity, externality, inequity, complexity, spatial dispersion, temporal
stability, perceived urgency and amenability to treatment. If we have five layers in
our road safety target hierarchy, and we have different indicators for each layer, it is
obvious that we need to combine and simplify information in order to help us
interpret. This is a good reason for the wish to capture all relevant pieces of
information in a composite index.

When benchmarking the safety performance of countries we have interests in
monitoring and understanding, if not explaining, progress in road safety.
Furthermore, we would also like to answer questions about how to adapt these
findings for other countries in order to enable countries to learn from each other.
And one step further: how to learn from outstanding practices in those countries that
are considered to be 'best-in-class'. When benchmarking, it is not necessary to
compare many countries, but it is a prerequisite to identify a 'best-in-class'. To this
end the countries must be grouped into classes, which gives rise to the question
how this must be done for this purpose. Can the grouping be carried out based on
2
road safety outcome indicators (final or intermediate), on policy output or policy input
indicators, or on the structural and cultural background of countries?
1.2. Aim of the study
It is obvious that there is not much history and experience in using indicators for
road safety. This is even more so the case for using a composite index. We
somewhat lag behind other fields. Examples of such indicators are known in other
domains such as the Human Development Index, which reflects life expectancy,
education level and living standards in a country, and is used by the United Nations
for the estimation of progress and annual country comparisons; the Environmental
Sustainability Index which is used by the World Economic Forum; or the Overall
Health System Index used by the World Health Organisation (WHO).

So far, we have used simple indicators in road safety like the number of people
killed in a road crash as the one single indicator. Sometimes (serious) injuries are
analysed additionally. When comparing countries we went one step further by
making indicators comparable by normalising or standardizing them, for example by
taking into account the size of a country, the number of inhabitants, motorization,
etc. In order to achieve a generally accepted way of normalization or standard-
ization, such as proposed by Trinca (Trinca et al., 1988), in SUNflower we used the
indicators personal safety (number of fatalities divided by the number of inhabitants)
and traffic safety (fatalities divided by the number of motorized vehicles). These
indicators, however, also raised some questions: which indicator is the best one,
can they be replaced by one another, etc. It is considered worthwhile to proceed
with this discussion and to investigate whether a composite index serves our goals
of making a more comprehensive comparison and of benchmarking road safety
performances between countries, more than simple indicators would.

Another interesting question deals with comparing programmes and performances
at a sub-national level, as was demonstrated in the SUNflower+6 study. In this study
we not only compared Greece, Portugal and Spain (Hayes et al., 2005), but we
could also compare a Spanish region (Catalonia) with Spain, and both other
countries. Sub-national comparisons open the possibility to take 'structure and
culture' (e.g. spatial and demographic factors, organizational and cultural factors)
into account to a larger extent.

Although it can be stated that working with indicators has not only advantages, the
anticipated benefits are considered appealing enough to study road safety indicators
in more detail. Based on the experiences with practical applications, this study
discusses the pros and cons of using indicators for road safety and policy making,
both at a national and a sub-national level.

The aim of this study is to develop a knowledge-based framework for comprehen-
sive benchmarking of road safety performances and developments of a country or of
other sub-national jurisdictions.
1.3. The SUNflower approach
As an introduction to the SUNflower approach we will refer to earlier studies and
publications (Koornstra et al., 2002; and Wegman et al., 2005). This study can be
3
seen as a logical follow-up of earlier work. Some of the SUNflower+6 study's
recommendations for further research can be used to illustrate this.

"We recommend the Commission to focus specifically on three major data issues,
exposure data, information on safety performance indicators and information on
severely injured road users.

In addition, we recommend to develop standards for the definition of such indicators
and for data collection procedures, in order to achieve unambiguous European data
that can be compared at the European level. Another challenging task is to soundly
quantify the relationships between particular levels of the road safety pyramid,
especially between the levels of indicators and outcomes, and to introduce the
methods on how to use this knowledge for the prediction and monitoring of road
safety outcomes at the country level.

Further knowledge development should be stimulated in order to assure that the
footprint gives a valid and reliable representation of countries' road safety
performances, now and in the future.

Finally, a prototype of a benchmark system has been developed; the data template
used in this project should be improved. We recommend that a European standard
will be developed of such a safety template, to be used in all European (Union)
countries. We further recommend to develop the existing and already working
prototype of a benchmark system into a user friendly final format for use with the
safety template."

The so-called footprint study discusses these recommendations in somewhat more
detail (Morsink et al., 2005).

After consultation with the European Commission and the SafetyNet Steering
Group, it was decided to explore possibilities to integrate these recommendations in
the SafetyNet project. In 2007, a SafetyNet-SUNflower workshop discussed how this
could be done.

The main conclusions from this workshop can be summarized as follows:

It was concluded that SUNflower can be of great added value to SafetyNet as a
valuable tool for benchmarking of the safety performance of countries. Although the
focus during the workshop was on the pyramid structure, SUNflower entails a lot
more than just the pyramid: it is more than a benchmarking instrument; it improves
our understanding of developments and consequently contributes to better
policymaking.

The pyramid shape gives the model a stable basis. The costs are at the top: after
all, we want to reduce the costs of crashes to society. However, there are some
important issues concerning the pyramid structure that need our attention on the
short term. Definitions are needed for mobility and exposure. When are they internal
in the pyramid and when are they external factors? What disaggregation levels for
the third dimension of the pyramid are most appropriate? Last but not least, it was
remarked, there is more work to be done in describing or developing clear indicators
for the different levels of the pyramid and for the links between them.

4
Finally, while ERSO is growing in importance, we need to establish its position in the
world, and providing a sound methodological framework can help in reaching this.

We can use the experiences and the results in the several SUNflower projects until
now as a solid basis for further enhancing our methodological framework for
benchmarking road safety performances. In addition, we can use certain results of
the SafetyNet project, especially those from WP1 (accident data and analysis), WP2
(exposure), and WP3 (safety performance indicators). The SUNflower approach
uses a so-called target hierarchy as presented in Figure 1.1.



Figure 1.1. A target hierarchy for road safety (Koornstra et al., 2002; LTSA, 2000).


Using this target hierarchy generated quite some support over the years and has
many followers. But this approach also raised discussion. First of all, the pyramid
was considered to be a too simplistic model of a far more complex reality. This may
be the case for all models. But if we do not start with a simple model to deepen our
understanding of this complex reality, we will most probably never be able to take
further steps.

Four items were subjects of discussion in the past:
the system boundaries of the pyramid and the definition and characteristics of the
external factors influencing the processes in the pyramid (Section 1.3.1);
how to define both bottom layers of the pyramid: Safety measures and
programmes and Structure and culture (Sections 1.3.2 and 1.3.3);
the vertical relationships between the different layers (Section 1.3.4);
the appropriateness of the top level of the pyramid: social costs (Section 1.3.5).
1.3.1. System boundaries and external influences
The first question that needs to be answered is which boundaries must be used in
the target hierarchy, which factors are part of the road safety system, and which
factors must be considered as separate from that system an why. Human activities
like traffic participation, vehicle choice and route choice are considered to be part of
the road safety system. This is the case even although these choices are hardly or
not at all made from a road safety perspective but are much more based on the
5
availability of a vehicle or a road, and an assessment of the cost and time a journey
will take. The reason is that the exposure to risk is a crucial factor within road safety,
like, for example, human activities are part of the Pressure-State-Response model in
the field of Environmental Performance Indicators; see Adriaanse (1993). The same
argumentation holds true for the (safety quality of the) design and the lay-out of our
road infrastructure and also for vehicles. It needs not be said that neither the road
design nor the use of road and vehicle can only be understood and influenced from
a road safety perspective, but also from a contribution to economic developments,
spatial planning, environmental effects etc.

In many countries a discussion has arisen about the road safety benefits that can
still be achieved being only relatively small, and why only specific road safety
measures should be applied. After all, there is a good driver training, there is a fair
amount of enforcement, road safety is included in the guidelines for road design to a
satisfactory extent, etc. This makes an answer interesting to the question if 'win-win'
situations can be achieved by strategic alliances with other social issues. Examples
could be health care, developments in giving society a more sustainable character,
social developments, etc. The search for win-win situations could have as a result
that the system boundaries for road safety become wider, but this needs to be
judged for each individual case.

For a full and correct picture of indicators at all levels of the pyramid we need to pay
attention to developments which affect the quality of measuring these indicators. As
an example, underreporting of crashes is a major problem in almost all countries.
The less severe the consequence of a crash, the higher the chance of not reporting
the crash to end by the police (Derriks & Mak, 2007). Also when it comes to
measuring safety performance indicators (Hakkert et al., 2007) large steps are still
needed to arrive at high quality comparable results.

The conclusion seems to be that there are no correct or incorrect system
boundaries, but that these boundaries are somewhat flexible. But it must be
recommended to investigate how to set the system boundaries for each problem
definition.
1.3.2. Road safety management: safety measures and programmes
The layer called Safety measures and programmes is an essential layer in the
pyramid, because it is by implementing effective (and efficient) measures and
programmes that we try to reduce the negative consequences of road crashes for
society, the so-called outcomes. All our efforts for a better understanding of road
safety and a better insight in measures and programmes are irrelevant if it cannot be
used to design and implement more and better measures and programmes.

So far, road safety activities studied in SUNflower have covered a long period of
time (from 1970 onward). However, it turned out that these interventions were
seldom a well-documented. This may be part of the explanation why road safety
developments could not be described and explained very well. Later we added more
general information to the evaluation items for policy documents and for effective
policy implementation (Wegman, 2004). This general checklist has not yet been
translated for road safety management purposes. However, the World Bank took the
first steps with its so-called country capacity reviews, which have in the meantime
been carried out for several countries, and the results of which serve as a basis for
further investments in improving road safety.
6

Recently it has been argued that it would be better to divide this layer of 'safety
measures and programmes' into two components: institutional road safety
management functions, a number of generic characteristics that allow for the proper
design and implementation of effective interventions, and the interventions
themselves (Bliss & Breen, to be published). This concept of 'managing for results'
will be further discussed as part of Section 2.4.
1.3.3. Structure and culture
The lowest layer/level of the pyramid, called Structure and culture has not yet been
very well defined in the SUNflower approach. SUNflower added an extra layer to the
model as developed in New Zealand (LTSA, 2000). The reasons were twofold:
It gives an essential background for all the observations and indicators at a
higher level of the pyramid. Progress in road safety could perhaps not be fully
understood or even be misinterpreted by not knowing or ignoring these
backgrounds.
It is not easy to transfer findings of benchmarking and to learn from experiences
and results abroad without having a clear picture of the setting in which these
results have been made or the changes were measured.

The SUNflower approach has been criticized for not fully recognizing the role of
spatial and demographic factors (IIHS, 2006) and organizational and cultural factors
(Delorme & Lassarre, 2005) in influencing casualty trends. In fact, the SUNflower
approach, and the pyramid on which it is based, include both these groups of
factors. However, it is fair to say that the influence of these factors on the work to
date has been explored to a much lesser extent than the data on more directly
safety related policies, such as accident outcomes, safety performance indicators
and policy inputs. Analyses at sub-national level provide one opportunity to explore
some aspects of these issues further (see Figure 1.2).

In the Structure part of the bottom layer two dimensions are distinguished: physical
structure and operational (functional) structure.

The physical structure of a country can be described by numerous factors that can
be defined as specific long-term conditions contributing to different road safety
outcomes. They are typically not, or at least not only, amenable to interventions by
conventional road safety policies. Moreover, they are typically modifiable by more
general policies, in a long term only. The two groups of structural factors can be
distinguished by their amenability to interventions in time: 1) Stationary factors not
changing in time (e.g. geographic and climate conditions) and 2) Tractable
(dynamic) factors subject to evolutions or changes in time (e.g. demography, road
topology, and urbanization).

7

Figure 1.2. Aggregated road safety pyramid (adapted from Eksler, to be published).

SOCIAL, ECONOMICAL, DEMOCRATIC STRUCTURE

PHYSICAL STRUCTURE

SOCIAL

COSTS

NUMBER KILLED

AND INJURED

SAFETY PERFORMANCE

INDICATORS

ACTORS

INSTITUTIONS

ORGANIZATION

VALUES

NORMS

ACTIONS

, MEASURES











TERRITORY

ROAD NETWORK

ROAD TRAFFIC



Stationary factors are of a physical nature and are beyond the influence of any
policy interventions, while tractable factors often have socio-demographical-
economical character and in a longer term can be influenced by targeted measures.

Stationary factors Tractable factors
Relief
Climate
Settlement geography
Demography
Urbanization
Road network topology
Social deprivation
Economical performance
Modal split
Table 1.1. Overview of some structural factors in road safety.


The list of structural factors presented in Table 1.1 is not exhaustive and many other
structural factors could be added. Also, it can be argued that many of the structural
factors are subject to adaptation processes, such as climate conditions, implying
different road infrastructures, and vehicle properties, but also different driving skills.
Therefore, these differences tend to have no real impact on compared road safety
outcomes indicators. However, there are other factors that indirectly have such a
strong impact on road safety outcomes, that they should not be omitted in any
relevant comparison of road safety performance. Structural factors describing the
settlement (and road network) structure have been identified as the strongest
determinative factors of road safety outcomes at a regional level (Eksler et al.,
2008a). They have such a strong impact on the speed driven by motorized vehicles
that they have a direct relation with road safety outcomes. Both qualified and
quantified indicators, i.e. typological or empirical indicators, could be considered.

The operational structure refers to the organization of and arrangements between all
potential actors involved in policy making. Therefore, this is where the manner is
8
discussed in which society uses institutions to try and solve social problems; road
safety in this case. The World report on road traffic injury prevention (Peden et al.,
2004) gives a good illustration of the numbers and variety of the different actors
(Figure 1.3).

ROAD INJURY
PREVENTION
POLICY
GOVERNMENT &
LEGISLATIVE BODIES
USERS / CITIZENS MEDIA
INDUSTRY PROFESSIONALS
POLICE
NGOs, SPECIAL
INTEREST GROUPS
ROAD INJURY
PREVENTION
POLICY
GOVERNMENT &
LEGISLATIVE BODIES
USERS / CITIZENS MEDIA
INDUSTRY PROFESSIONALS
POLICE
NGOs, SPECIAL
INTEREST GROUPS
GOVERNMENT &
LEGISLATIVE BODIES
USERS / CITIZENS MEDIA
INDUSTRY PROFESSIONALS
POLICE
NGOs, SPECIAL
INTEREST GROUPS

Figure 1.3. Overview of different key stakeholders in road safety policy (source:
Peden et al., 2004).


Somehow, agreements will have to be made between the various actors about their
contribution to road safety improvement. If this cooperation is not well coordinated,
loss of both quality and efficiency will be the result. An added complication is that
such losses are not easily indicated. The solution to this problem could be to ask
organizations to commit themselves to verifiable performances and next to create a
system that makes them accountable for these performances. If different actors are
expected to produce policy performances at the same time and if the joint total of
these performances is more than the sum of it parts, inadequate cooperation results
in (unnecessary) loss of effectiveness.

The government's role is especially important in this, or rather the roles of the
different layers of government. It must be said here that the government is not only
committed to keep to its own agreements and to deliver a good product at the lowest
possible cost. The government has to 'deliver' in a political context where political
rationalities are important and play a serious part alongside the scientific
rationalities.

As we remarked earlier, it is not yet customary to thoroughly document the policy
efforts for road safety improvements in terms of deliveries, products and costs. This
complicates the scientific evaluation of implemented policies that was pursued and
hence also the giving a proper answer to the question if the delivery has been
adequate for achieving the target that was set.

Culture consists of values and norms in their social sense. Values can be regarded
as assumptions upon which implementation can be based. Sets of consistent values
and measures together form the value system, which is subjective and varies across
people. Types of value include ethical/moral values, ideological, social and aesthetic
9
values and it may be argued that all of them have an influence on behavioural
attitudes of road users, which in turn will manifest itself in different road safety
outcomes. Values such as the value of a human life, respect for each others rights,
etc., are directly reflected in road safety provisions, such as those related to
reduction targets. Norms refer to the rules that are socially enforced. Social
sanctioning is what distinguishes them from values. They can be viewed as
reference standards, or statements that regulate behaviour and act as informal
social control. The most typical example is society's attitude towards drink-driving,
which differs significantly between countries.

For road safety this is reflected in the way society deals with the consequences of
the lack of road safety, to what extent these consequences are considered to be
unavoidable, and the degree of social and political interest in eliminating or at least
modifying these consequences. This is about road safety culture and this culture
partly decides the political, governmental and social reactions to traffic risks (see
also AAA, 2007). What role does a government see for itself in reducing risks in
society, and where is the boundary between citizens', respectively road users',
responsibility and collective responsibility? And how do political priorities translate
this collective responsibility? And to which extent will road safety measures be
accepted, especially if they limit individual freedom? But the cultural element can
also be seen as the way in which a society, and politics in particular, deals with
setting concrete goals (a quantitative road safety target) and the reaching or failing
to reach such a target. Undoubtedly countries differ, but previously a European
study of drivers' attitudes has taught us that there is also a reasonable amount of
similarity between countries, as has been illustrated in several SARTRE studies
(1991,1996, 2004).

However, it is certain that when comparing countries differences in 'structure and
culture' have an effect on the size and nature of road safety problems, but also
influence the possibilities to reduce the problem effectively and efficiently. This
presents an important theme for future research.
1.3.4. Vertical relationships between the different layers
The presentation in the shape of a pyramid with different layers could give the
impression that the layers are relatively unconnected. Nothing is further from the
truth, as has been made clear in earlier SUNflower publications. In fact, the pyramid
has three, if not four, dimensions. Two of these dimensions are not visible in Figures
1.1 and 1.2. The first is the dimension time. The pyramid's indicators can be read
periodically and this way trends can be studied. The second dimension is that
indicators can be read not only for a country in its entirity, but also for (parts of the
problem): regions, modes of transport, road types, age groups, etc. This can be
visualized by not using a triangle like in Figures 1.1 and 1.2, but by adding a third
dimension, thus creating a pyramid.

The two remaining dimensions can also be made visible in the two-dimensional
plane, the triangle. The horizontal dimension indicates that the number of
observations decreases for the top three layers while approaching the top. The
pyramid's layers are stacked logically. This enables a top-down approach:
understanding developments at the top and explaining them using developments at
the bottom. It is also possible to make changes at the bottom and investigate to
what extent they cause changes at the top.

10
The relations between indicators at different layers are very important and must be,
conceptually seen, causal for the top four layers. Without these causal relations the
pyramid is even meaningless. We will use one example as an illustration. Policy
interventions will first need to have an effect at the level of the intermediate variables
(SPIs) before it can be made credible that the interventions have an effect on
crashes and risks. Alcohol legislation will first have to result in fewer alcohol-related
crashes and fewer alcohol-related casualties.

It goes without saying that our knowledge is not good enough to link indicators of the
different layers in a causal way. Although we lack for evidence-based information,
we may use the judgement of road safety experts to overcome this drawback. This
lack of 'evidence-based' information can be considered as a good incentive to guide
further research.

It can be observed that interventions are increasingly composite interventions: it is
not just new legislation, but they also include the public information about the new
legislation and its enforcement. It can become even more complicated when it is
attempted to discourage the use of alcohol in a society, one of the reasons being to
reduce traffic participation under the influence of alcohol. In addition, determining
the effects of interventions becomes increasingly difficult if they are more widely
spread over time and place. This presents a heavy task for the methodology of
evaluation research.

Summarizing, the idea behind the pyramid's layers is the continuous attempt to
define a causal relation between the top four layers which can be seen as the core
of the SUNflower approach.
1.3.5. Social costs
Until now, SUNflower has paid hardly any attention to the social consequences of
road crashes, more in particular to those consequences that can be expressed in
monetary units. There are good reasons to initiate this (SWOV Fact sheet Road
crash costs, 2007) and therefore Social costs have been added to the pyramid as its
top layer. In the first place, this information is useful for comparing road safety policy
with other policy areas. These can be other sectors within traffic and transport or
outside, for instance environmental care, public health or other safety issues.
Secondly, information about the costs of road crashes is used in cost-benefit
analyses (SWOV Fact sheet Cost-benefit analyses of road safety measures, 2008).
Social costs estimates can be used for setting policy priorities.

There is another good reason for adding Social costs as a top layer to the pyramid,
rather than ending with 'Numbers of killed and injured'. This will be illustrated with an
example. Assuming that the development of the number of fatalities is not exactly
equal to the development of the number of injuries, which conclusion is to be
drawn? This is not a hypothetical question, but a reality in many countries. This
situation requires a method of adding up fatalities and injuries. However, fatalities
are considered to be more severe than injuries, and, moreover, there are major
differences between the severity of injuries: from lifelong disability to a bleeding
thumb. To take this into account the health sector has developed indicators such as
QALY (Quality Adjusted Life Years) and DALY (Disability Adjusted Life Years)
(Sassi, 2006). Injuries scales such as the Injury Severity Scale and the Abbreviated
Injury Scale have also been developed. All these scales help us attaching a value to
different consequences of road crashes. The SafetyNet project made a major
11
contribution to the comparability of road crash injury data between European
countries.

SUNflower uses the term Social costs, sometimes the term Road crash costs is
used, or the term socio-economic costs. All three terms cover five main categories:
medical costs;
production loss;
quality of life loss;
material costs;
settlement costs.

Estimates of these costs have been made for several European countries (Elvik,
2000). They vary form 1.3% to 3.2% of the Gross Domestic Product (an average of
2.1%). This type of information allows us to make comparisons with other sectors in
society. It also enables rational prioritization of policy actions based on cost-benefit
analyses.

Therefore, we have every reason to place the consequences of road crashes at the
top of our pyramid and we have to develop procedures/methods to use these
estimates properly in benchmarking the safety performance of countries.
1.4. Structure of the report
Chapter 2 discusses the concept of road safety benchmarking and the use of road
safety indicators for that purpose. It is argued that it might be an excellent idea to
capture the complex phenomenon of road safety in some simple indicators, if not in
one single composite index. Simplification, quantification and communication are the
key words here. However, indicators should be accepted by road safety researchers
and professionals, as well as by policy makers. This chapter hopes to gain that
support. Different types of benchmarking are distinguished and introduced.

Chapter 3 makes a first proposal for a composite index for road safety performance.
It is our ambition to include the different layers of the road safety pyramid in such a
composite index. The proposed layers are policy performance indicators (safety
programmes), road safety performance indicators (killed and injured) and
implementation performance indicators (limited to a set of measurable safety
performance indicators). The aim of this composite index is to enable ranking
countries in accordance with their safety performance.

The concept of benchmarking not only addresses ranking the safety performances
of countries; this ranking is only a step towards 'identifying, understanding and
adapting outstanding practices from the countries which are considered to be 'best-
in-class'. Therefore, this concept requires defining classes and identifying criteria
that can be used to form different classes. Therefore Chapter 4 is dedicated to the
problem of how to group European countries and it answers the question whether it
is wise to form different classes or whether to consider all European countries to be
pupils in the same class.

Chapter 5 deals with road safety developments and seeks how to describe and
analyse these developments best. Trends for fatality risks and rates are presented
for individual countries and the results are compared. For example, the results of the
grouping of countries from Chapter 4 are used to illustrate the potential conclusions
12
that can be based on the modern techniques of time series analyses. The same
techniques are applied to studying disaggregate data for age groups, traffic modes
and road types.

Chapter 6 starts with the observation that SUNflower analyses to date have
focussed on national comparisons. These analyses can also be made for the
comparison of programmes and performances at a sub-national level. This allows us
to take into account structural (spatial, demographical, economical, political) and
cultural differences/variation and to gain better understanding of their importance.
Sub-national analysis can be made at a regional level and to compare the safety
programmes and performances of cities. Avenues for further work are presented.

The final chapter, Chapter 7, contains conclusions and recommendations. It
summarizes the main findings and uses them as a basis to draw conclusions and
make recommendations for next steps.


13
2. Benchmarking by using road safety indicators
2.1. Benchmarking road safety performances
Basically, the essence of (international) cooperation is learning from each other.
This learning should be targeted at a better understanding of the subject involved, in
our case road safety. The learning includes subjects such as monitoring and
explaining road safety developments, and gaining good insights in the impacts of
interventions (in the causal relationships between interventions and impacts on road
safety, in the active ingredients of interventions and in the dose-response
relationship) as a basis to speed up improvements in road safety in one's own
country or jurisdiction.

Benchmarking is a process in which countries or jurisdictions (states, provinces,
'lnder', etc.) evaluate various aspects of their performance in relation to other, and
so-called 'best-in-class' practices. The benchmark results provide countries or
jurisdictions with information from others that can be used as a basis for developing
measures and programmes to increase their own performance. From here on we
will only mention country or countries in this chapter, but the sub-national level,
consisting of regions and jurisdictions, is also included.

Benchmarking consists of the following core activities: identifying the key
components of a road safety performance, identifying with whom to compare (other
countries/jurisdictions and 'best-in-class'), constructing indicators for meaningful
comparisons, determining and understanding gaps in performances, and, finally,
establishing future attainable performances. It is attractive to speak about a
benchmark cycle (Figure 2.1) and to carry out benchmarking at regular intervals, to
monitor progress made and to evaluate the results of interventions.

Comparing
Developing (set
of) indicators and
with whom
and how
Comparing
and analysing
with best in
class
instruments
ermining
Identifying
ntation
Evaluation
Det
and under-
standing gaps in
performance
Impleme
potential
improvements
Comparing
instruments
ermining
Identifying
ntation
Evaluation
Comparing
instruments
ermining
Identifying
ntation
Evaluation
Developing (set
of) indicators and
Developing (set
of) indicators and
with whom
and how
Comparing
and analysing
with best in
class
with whom
and how
Comparing
and analysing
with best in
class
Det Det
and under-
standing gaps in
performance
Impleme
and under-
standing gaps in
performance
Impleme
potential
improvements
potential
improvements

Figure 2.1. The benchmarking process in seven steps.

14
Benchmarking is based on learning from others, rather than developing new and
improved approaches. Although helpful, benchmarking should never be the primary
strategy for improvement. However, it can lay an important basis for a good
strategy.

Two important tasks can be identified for this process:
defining the key components of a road safety performance and investigating if
and how these key components can be brought together in a composite index, a
uman Development Index) and an
indicator.
field of road safety has
cation, to
t al., 2004), in
pment Index, which reflects life expectancy, education level and living
es a multiple score of standardized key indicators.
The indicators can be compared with meaningful references.
road safety performance index;
finding a meaningful 'reference' (best-in-class) and defining procedures for
identifying such a meaningful reference.

In the literature and in practice two words are regularly used for ranking
performances: an index (e.g. Dow Jones Index, H

The number of indicators which is suggested for use in the
been growing rapidly, especially over the last decade (e.g. ETSC, 2001; Wegman et
al., 2005; Hakkert et al., 2007). Today, recognizing the complex character of the
road safety phenomenon, more and more indicators are used with the intention of
measuring the factors leading to accidents, identifying conditions which are
associated with increased accident/injury risks, and detailing the structure of traffic
injury patterns, whereas the traditional approach considered the safety outcomes
mostly in terms of fatalities per head of population, vehicle fleet or exposure.

Because the word 'indicator' is so heavily used in road safety already, we decided to
work with the word 'index': a road safety performance index. Because this index is a
combination of several performance indicators, we introduce the term composite
dex in this study. Perhaps it will be helpful, for reasons of easy communi in
call this performance index the SUNflower index.
2.2. Performance indicators for road safety
Road safety is steadily developing into a major policy area (Pe
which safety performance indicators should serve as su
den e
pportive tools for
policymakers. In comparing the safety achievements of countries there is a need to
reduce the dimensions of the problem and to be able to work with a composite index
that can express all the relevant components in a concise and comprehensive way.
Examples of such indicators are known in other domains such as the Human
evelo D
standards in each country, and is used by the United Nations for the estimation of
progress and annual country comparisons, the Environmental Sustainability Index
which is used by the World Economic Forum, the Overall Health System Index used
by the World Health Organisation (WHO), and others (Nardo et al., 2005).

In SUNflower+6 the concept of road safety footprints was developed. A road safety
footprint of a country was described by Morsink (2005) as a representation of the
road safety status of a country. Three components of this footprint were considered
to be essential:
The footprint giv

15
The indicators are expressed as a snapshot in time, and as a past picture over
time.
A road safety footprint draws a full picture of all impacts of road safety and their
most relevant underlying elements and processes for which causal relationships
exist.

The SUNflower+6 final report suggested that the road safety footprint of a country
ideally is a composition of suitable indictors at all levels of the so-called SUNflower
yramid and for all components of the traffic system. Morsink et al. (2005)
footprint
oncept was introduced. Ecological footprint analysis compares human demand on
Adriaanse (1993). Demand
than supply in order to create a sustainable situation. For road safety
helpful for road safety.
he first chapter already introduced the main aim of this study as to develop a
to leave out the concept of footprint, although both approaches
re rather similar and work in the same direction.
p
developed two schemes: a detailed footprint scheme and a summary footprint
scheme. Both schemes intend to give an overview of indicators at all layers of the
pyramid and make proposals how to compare indicators with a reference. It was
experienced that even the summary schemes carry too many pieces of information
to be understood easily and were not considered as attractive enough to be helpful
for policy makers.

The concept footprint received a boost a decade ago when the ecological
c
nature with the biosphere's ability, or it regenerates resources and provide services
(Wackernagel & Rees, 1996). This ecological footprint approach tries to include
environmental externalities in decision-making. For products and services the
assessment of the environmental impact of a given product or service throughout its
lifespan can be expressed by using the concept of life cycle assessment; see for
example ISO 14040 (ISO, 2006).

The (ecological) footprint concept is basically a comparison of demand and supply,
as the Pressure-state-response model developed by
should be less
this is not the case. We cannot identify a safe supply level other than zero
fatalities/injuries, and we understand well that reaching zero is unachievable under
prevailing conditions. Perhaps the lack of a clear-cut and acceptable supply level in
road safety complicates an easy acceptance of this metaphor too much and seems
to support the association with footprint as not very

The concept of comparing performances and, one step further, that of benchmarking
performances seems to be a more appropriate approach for road safety.
Benchmarking was originally developed and used in the private sector to compare
the performances of individual companies as a tool for improving their operations.
Benchmarking tries to provide an objective way of measuring performances against
a meaningful reference (in the private sector: the competitor). This meaningful
reference is sometimes described as the performance of the 'best-in-class'. This
comparison is usually made with the aim of increasing some aspects of its own
performance, in other words to learn from each other.

T
benchmark cycle for international comparisons for road safety performances. In
order to prevent misunderstandings we propose to work with the concept of
benchmarking, and
a

Three main functions of indicators have been made clear by Adriaanse (1993) in his
attempts to build indicators to be used in environmental policies: simplification,
quantification and communication. This implies that defining indicators should be
16
directed by which intentions must be satisfied when using indicators. Basically, by
using indicators we try to capture complex phenomena in relative simple terms and
in doing this, we run the risk of losing relevant information or insights. Nevertheless,
according to Adriaanse, indicators generally use simplification to make complex
henomena quantifiable in such a manner that communication is either enabled or
dicator' and
roduct indicator' are also used. In SUNflower we distinguish final outcomes
or a meaningful comparison of countries, numbers of people killed or injured are
The second type of indicator indicates the quality of the implementation of road
Implementation performance, in general, can deal with different components of
egies, programmes, resources, coordi-
ation, institutional settings, etc.) and the quality of action plans and individual
p
promoted. Furthermore, these indicators can be used to compare countries, to rank
them and to benchmark them.

In SUNflowerNext three types of indicators are distinguished. These three indicators
together could, and in our view should be combined to form one composite index
(see Section 2.3).

The first indicator captures the quality of road safety in a country and has been
named Road safety performance indicator. The terms 'outcome in
'p
(numbers of killed and injured), intermediate outcomes (such as the safety
performance indicator), and social costs. In this we follow the line chosen in the New
Zealand model (LTSA, 2000) and later in SUNflower (Koornstra et al., 2002).
Therefore, we use the top three layers of the pyramid (see Figure 1.1) to indicate a
country's road safety performance.

F
typically 'normalized', which results in fatality rates, e.g. fatalities per inhabitant,
vehicle, or kilometre travelled. In addition, more vulnerable groups of road users like
pedestrians, cyclists, and motorized two-wheelers may specifically be considered.
Based on the number of people killed and injured and the consequences of road
accidents it is possible to express the socio-economic burden imposed on societies
by road accidents. These costs enable a comparison of the consequences of road
crashes with other threats to public health (Peden et al., 2004) or with other
investment priorities in a country or jurisdiction (Elvik & Veisten, 2005).

safety policies and has therefore been named Implementation performance
indicator. For this implementation quality indicator the term 'process indicator' can
also be used. Basically, this indicator follows a vertical line in the pyramid linking
'safety measures and programmes', safety performance indicators, and the numbers
of people killed and injured.

causal relationships between the different layers of the safety pyramid, such as
between the policy context ('structure and culture') and the road safety policies;
between the policy changes ('safety measures and programmes') and the changes
in performance indicators; between the changes in safety performance indicators
and changes in the number of casualties, et cetera. In this context, quite some
progress was attained in the development of safety performance indicators within
the SafetyNet project. However, the possibilities for systematic measurement of all
these relationships are still limited and need further research.
The third type of indicator indicates the quality of response in policy documents to
improve road safety and has been named Policy performance indicator. It has two
components: the quality of conditions (strat
n
17
(counter)measures in the perspective of ambitions of countries as expressed in road
safety targets
Policy performance deals with the quality of road safety strategy, more specifically
with the quality of road safety plans and with the conditions introduced for the actual
implementation of road safety measures and programmes, e.g. institutional
arrangements, budget, quality of professionals, application of evidence-based
knowledge, sound analysis and diagnosis of road safety problems, vertical
cooperation between different tiers of government, etc. Reports by OECD (2002),
ETSC (2006), Bliss & Breen (to be published) summarize the demands for the
effective development and implementation of national road safety policies.

Hence, three types of indicator, Road safety performance indicator, Implementation
performance indicator and Policy performances indicator can be included in the
target hierarchy of the pyramid (Figure 2.2).

Structure and culture
Safety measures and programmes
Policy performance
indicator
Safety perform
Number killed and injured
Implementation performance
indicator
indicator
ance indicators
Road safety performance
(Policy context)
Social costs
Structure and culture
ance indicators
Road safety performance
(Policy context)
Social costs
Safety measures and programmes
Policy performance
indicator
Safety perform
Number killed and injured
Implementation performance
indicator
indicator
Structure and culture
ance indicators
Road safety performance
(Policy context)
Social costs
Safety perform
Number killed and injured
Implementation performance
indicator
indicator
Safety measures and programmes
Policy performance
indicator
Figure 2.2. Three types of indicator combined, forming the performance index for
d safety. roa
2.3. Road safety: towards a composite performance index
The target hierarchy for road safety, as presented and used in earlier SUNflower
benchmarking. For several reasons, it is attractive to co
apted from Saisana &
r
studies, and in the SafetyNet project, contains several building blocks to be used in
mbine all information in one
indicator, a so-called composite index.

The main pros and cons of using composite indices are ad
a antola (2002) by OECD/JRC (2008). T

Pros
Composite indices:
can summarize complex, multi-dimensional realities with a view to supporting
decision-makers;
are easier to interpret than a battery of many separate indicators;
can assess progress of countries over time;
18
reduce the visible size of a set of indicators without dropping the underlying

facilitate communication with general public (i.e. citizens, media, etc.) and
increase the difficulty of
n, if the construction process is not transparent;
es if dimensions of performance that are difficult
e is extremely useful in garnering media interest and hence
d to imagine that debate on the use of composite indicators will ever be
ind for constructing and using composite
The
pe ss, sustainable development,
information base;
thus make it possible to include more information within the existing size limit;
place issues of country performance and progress at the centre of the policy
arena;

promote accountability;
enable users to compare complex dimensions effectively.

Cons
Composite indices:
may send misleading policy messages if they are poorly constructed or
misinterpreted;
may invite simplistic policy conclusions;
may be misused, e.g., to support a desired policy, if the construction process is
not transparent and/or lacks sound statistical or conceptual principles;
the selection of indicators and weights could be the subject of political dispute;
may disguise serious failings in some dimensions and
identifying proper remedial actio
may lead to inappropriate polici
to measure are ignored.

Nardo et al. (2005) quote two statements on the pros and cons of using composite
indices and these quotations illustrate these pros and cons rather well.

"The aggregators believe there are two major reasons that there is value in
combining indicators in some manner to produce a bottom line. They believe that
such a summary statistic can indeed capture reality and is meaningful, and that
tressing the bottom lin s
the attention of policy makers. The second school, the non-aggregators, believes
one should stop once an appropriate set of indicators has been created and not go
the further step of producing a composite index. Their key objection to aggregation
is what they see as the arbitrary nature of the weighting process by which the
variables are combined." (Sharpe, 2004, in Nardo, 2005.)

According to other commentators:

[] it is har "
settled. [] official statisticians may tend to resent composite indicators, whereby a
lot of work in data collection and editing is 'wasted' or 'hidden' behind a single
number of dubious significance. On the other hand, the temptation of stakeholders
and practitioners to summarize complex and sometime elusive processes (e.g.
stainability, su single market policy, etc.) into a single figure to benchmark country
performance for policy consumption seems likewise irresistible." (Saisana et al.,
05, in Nardo, 2005) 20

The OECD (Nardo et al., 2005) prepared a Handbook on constructing composite
icators with the aim to provide a guide
indices for policy-makers, academics, the media and other interested parties.
handbook is concerned with indicators which compare and rank country
rformance in areas such as industrial competitivene
globalization and innovation. The handbook contains a set of technical guidelines
19
that can help constructors of composite indices to improve the quality of their
tputs. ou
Th onstruction of composite
work - to provide a basis for selection and combination of single
indicators;
Robustness and sensitivity analysis should be undertaken to assess
ed
indic tify linkages through regressions;
Visualization composite indices can be visualized or presented in a number of
ere is a generous
mount of information and even the summary footprint scheme as developed by
is insufficient knowledge to

e handbook distinguishes the following steps in the c
indices, which are of relevance for a road safety composite index:
Theoretical frame
Data collection on the basis of their analytical soundness, measurability,
country coverage, relevance of the phenomenon being measured;
Multivariate analysis assess the suitability of the data set;
Imputation of missing data consideration should be given to different
approaches for imputing missing values;
Normalization indicators should be normalized to render them comparable;
Weighting and aggregation indicators should be aggregated and weighted
according to the underlying theoretical framework;
robustness of the composite index;
Links to other variables correlate the composite index with other publish
ators as well as to iden
different ways, which can influence their interpretation;
Back to the real data composite indices should be transparent and be able to
be decomposed.

Although it is evident that to construct and work with composite indices has not only
advantages, this fits the SUNflower approach rather well. Th
a
Morsink et al. (2005) requires a thorough comprehension of road safety and its
mutual relationships to understand this information in its full depth. This led to the
conclusion that a composite index for road safety needed to be developed in which
all components of the SUNflower pyramid should have a position. In addition,
attention was to be paid to the cons that are presented in this section.
2.4. Performance indicators for road safety policies and their
implementation
In recent years, there has been increasing attention for rational decision-making
about road safety; not in the last place caused by the fact that many countries use
quantitative targets. This indicates a policy and research based interest in the road
safety developments (will we or won't we make the target?) as well as in the factors
that can be an explanation for those developments (did the implemented policy
make a contribution to the developments that were observed?).

It is barely possible to give a coherent explanation for the road safety developments.
This is partly due to the complexity of the problem and the limitedness of our
knowledge. Many factors and developments are important and have an influence,
but at the same time we must establish that there still
obtain a clear picture of all the relations. In addition, much important information is
still lacking. But the growing attention for this area can be seen to lead to real
improvements.

20
From this perspective there are two areas that require our specific attention if we
wish to develop a SUNflower road safety performance index that allows all layers of
the pyramid to be visible: in the first place the quality of the intended road safety
policy (in Section 2.4.1), as can be found in the policy documents (Policy
performance indicators). But it is also important that we can measure the quality of
the policy implementation (in Section 2.4.2), which is indicated by the extent to
which the intended policy is implemented, and to what extent this policy
implementation affects the safety quality of vehicles, roads and human behaviour
(the safety performance indicators) and to what extent changes can affect the
numbers of crashes and victims. This means that here the quality of the causal
hain is under discussion. This quality was previously called the Implementation
few years much work has already been done on Road safety
erformance indicators. This includes setting the numbers of fatalities (and
1988), SUNflower
as also chosen this approach. In the discussion about the best standard SUNflower
ples are the idea of valuing injury severity or to find a way to
combine killed and seriously injured by taking the injury severity into account without
ent can be made by taking vulnerable
road users into account. The United Kingdom, for example, has decided to formulate
years, the second part of the Road safety performance
ediate outcomes, has been detailed by SPIs (safety
rs). Here, the SafetyNet project has made a very important
.4.1. Benchmarking policy performance
. Policy relevance;
4. Recogniza

c
performance indicator. Both these performances will briefly be discussed in this
chapter, and will be elaborated in the following chapter.

In the past
p
sometimes also the numbers injured) as the standards for international comparison.
This is often done by dividing these numbers by the size of the population (mortality
and morbidity), and/or by the number of vehicle kilometres, or, if that data is not
available, by the number of motorized vehicles. Following Trinca (
h
has chosen to present the possibilities as different, independent measures that are
all valuable. There is, however, increasing interest in detailing these general
measures. Exam
simply adding the numbers. A second refinem
a separate target for the safety of children.

During the last few
indicators, the interm
performance indicato
contribution.

The last part of the pyramid that must be mentioned is the bottom layer: structure
and culture. This is the foundation for road safety on which policy making and
implementation is based. But this bottom layer also indicates the possibilities and
limitations for policy. This layer also needs to be detailed further.

The remainder of the present chapter will discuss two indicators: de policy
performance indicator (Section 2.4.1) and the implementation performance indicator
(Section 2.4.2).
2
For Policy performance indicators four key points are identified as crucial by
Adriaanse (1993):
1. Quality aspects;
2. Sensitivity in time;
3
bility and clarity.
21
The first point o the methodology used.
This methodology must and scien-
tifically accep ce. The indicators should
enable composing temporal trends and identifying effects of medium-term and long-
term policy interventions. The developed methodology must be derived from the
main policy structure, paying attention to groups, not
only for a cou lly, recognizability
and clarity re an easy appeal in order for
them to be accepted by policy makers and scientists.

Policy perfor e 2.1 lists
the items tha
t this stage, uments, resulting in a score, is based on
to develop a standardized method-
rs at a regular basis.
concerns not only the quality of data but als
be clearly defined, accurately described, socially
table and, consequently, easy to reprodu
major policy themes and target
ntry as a whole, but also at a sub-national level. Fina
quire carefully designed indicators with
mance deals with the quality of a road safety strategy. Tabl
t are required for the evaluation of policy documents (Wegman, 2004).
the assessment of policy doc A
expert opinions. However, it is recommended
ology for this purpose and to collect data for these indicato

Evaluation items for policy documents
The political support of the document
The precision of the definition of goals/objects/targets
The use of valid causal theory (problem solution)
The available means (implementation + monitoring)
The reduced necessity of inter-organizational decisions
The sanctions/incentives for co-producers and target audience
The implementation priority for all stakeholders
The active support of stakeholders
Table 2.1. Evaluations items to measure the quality of policy
documents (Wegman, 2004).

2.4.2. Benchmarking implementation performances
Benchmarking implementation performance mainly involves the causal relations
between the different layers in the pyramid:
if and how policy changes (safety measures and programmes) affect safety
performance indicators;
l., 2002) introduced the causal chain
rformances of specific road safety interventions and their
tcomes, such as people
in New Zealand (LTSA,
o this model, titled Structure and culture.
was needed as a policy context for
if and how changes in safety performance indicators affect changes in the
number of casualties (killed and injured) and casualty rates;

The initial SUNflower study (Koornstra et a
between the policy pe
impact on Road safety performance indicators and final ou
as based on activities killed in a road crash. The study w
2000), but we decided to add a basic layer t
was argued that such a basic layer It
understanding (impacts) of road safety policy. The Structure-component addresses
topics such as how responsibilities between governmental layers are divided, and in
relation with this, how governmental budgets are organized, what is considered to
be a role for the private sector in road safety, what role NGO's play, etc. The culture
element concentrates on how a society and its citizens perceive the road safety
22
problem compared with, for example, the role of (motorized) traffic in our society and
its contribution to economical growth, welfare, environmental consequences, etc.
Also questions on how responsibilities are defined for individuals and the govern-
ment are relevant. We can also use the word 'safety culture' here (AAA, 2007).

Based on international literature on public policies Wegman (2004) lists the
circumstances that affect the implementation quality of the policy documents and
at are useful for the monitoring progress:
ially we
ystematically if and to what extent intended policy
t (Bliss & Breen, to be published) is interesting.
uires detailing in indicators. A
ore detailed definition is made in Chapter 6. Continuing this effort
in the
these

y illustrated above that a good understanding of the effect of an
actua
suffic
wheth
impro sible to not only consider the actual measure, but to pay
th
the economic/social/political environment;
the public support;
the progress of the implementation of policy documents;
the support of key stakeholders;
the quality of the 'delivery mechanisms'.

However, it must be conceded that this is really a new area within road safety; an
area, moreover, that is not yet very well-documented in many countries. If init
were almost exclusively working on understanding road safety from a further and
increasingly more in depth analysis of road traffic accidents, since only a few years
the interest for road safety performance indicators (SPIs) came into being. It was
mainly initiated by the Swedish professor Kare Rumar and was later developed by
ETSC (2001). SafetyNet (Workpackage 3) has further elaborated this concept
(Hakkert et al., 2007). Now a following step needs to be taken along this path:
investigating and documenting s
has been implemented, what relevant changes affecting road safety have occurred,
and, finally, whether these influences indeed resulted in changes in safety
performance indicators and the numbers of casualties.

In this context, the World Bank repor
This report is based on six recommendations in the WHO/World Bank report on road
safety (Peden et al., 2004). In the chapter entitled 'Managing for results' Bliss &
Breen discuss two layers that are important for effective policy implementation: first
a number of 'institutional management functions' need to be well organized and to
be embedded and operational. Within this bedding, effective and efficient measures
can then be implemented. The Bliss & Breen report can be regarded as a further
elaboration and detailing of the list that was formulated by Wegman in 2004.

The institutional management functions could also be fitted in the pyramid's bottom
layer (Structure and culture). But as we wish to reserve this layer for conditions and
developments that are suitable for wider application and can hardly be influenced
from the road safety perspective, the choice has been made to split the layer 'Safety
measures and programmes' in two parts: the management functions and the
concrete measures and actions.

In addition, the layer 'Structure and culture' also req
first attempt at a m
sequel to this study is recommended. Chapter 3 offers a first elaboration;
indicators fit within 'Structure'.
It has been clearl
l road safety measure is only possible if the bedding of such a measure is
iently well-known. This has two major consequences. If it is important to know
er a measure has been implemented well or whether there is room for
vement, it is sen
23
special attention to its bedding. Making explicit advance statements about a
meas
of an
gettin
conse from the other while the
easure itself, but also its bedding needs to be
taken
succe

iscuss seven management functions:
Re
ns agreed to achieve this ambition.
ips to achieve the desired
Le
heir interventions, and their related institutional management
functions to achieve the desired focus on results.
s to achieve the desired focus on results.
Promotion
terpretation to these seven functions from the SUNflower pyramid concept.
ure's expected output (and about its expected contribution to the improvement
SPI or the numbers of fatalities and injured) increases the possibilities of
g a thorough understanding of the effects of measures. There is also a second
quence: if the one country wants to learn
circumstances differ, not only the m
into account. Only then can insight be attained in the possibilities of
ssfully 'importing' a measure that has been effective elsewhere.
Bliss & Breen d
sults focus
The foremost and pivotal institutional management function which can be
interpreted as a pragmatic specification of its ambition to improve road safety
and the mea
Coordination
This concerns the orchestration and alignment of the interventions and other
related institutional management functions delivered by government partners
and related community and business partnersh
focus on results.
gislation
This concerns the legal instruments necessary for governance purposes to
specify the legitimate bounds of institutions, their responsibilities and
accountabilities, t
Funding and resource allocation
This concerns the financing of interventions and related institutional
management functions on a sustainable basis using rational evaluation
framework to allocate resource

This concerns the sustained communication of road safety as a core business

for government and society emphasizing the shared societal responsibility to
support the delivery of the interventions required to achieve the desired focus
on results.
Monitoring and evaluation
This concerns the systematic and ongoing measurement of road safety
outputs and outcomes (intermediate and final) and evaluation of interventions
in terms of achieving the desired focus on results
Research and development and knowledge transfer
This concerns the systematic and ongoing creation, codification, transfer and
application of knowledge that contributes to the improved efficiency and
effectiveness of the road safety management system to achieve the desired
focus on results.

Hence, the basic concept here is 'results focus' and then the different functions are
considered from this perspective. It is an attractive idea to try and attach an
in
Chapter 3 makes an initial attempt.

The Bliss & Breen report rather emphasizes the government's role in increasing
road safety and, indeed, there are no examples of a real and lasting progress
without the government having a (very) prominent role. Therefore, it is under-
standable that the WHO/World Bank report argues for a 'lead agency' as a drive
behind the initiatives for and the implementation of road safety policy. In countries
24
where insurance companies are part of the government, like in Victoria, Australia,
and in British Columbia in Canada, these companies are active precisely because
they are part of the government. But here it must not be underestimated that it is not
just the role of a central government that is important. This is evident in countries
with a federal form of government. But much policy is made by regional and local
governments and in and by local communities. Here the hidden demand for more
road safety, for example a safe route to school for children or dangerous road
section treatment, can fulfil an important role. Social organizations could, and
sometimes have an important role in this. And their role in a successful road safety
policy should really also be made more visible, and therefore be incorporated in
performance indicators. It is recommended to pay attention to this in the near future.

Finally we will need to address the individual responsibility of citizen and road user,
s possible to incorporate this in a performance indicator.
After all, he or she carries the responsibility to avoid accidents. But this road user
hoice architecture to nudge us in
eneficial directions without restricting freedom of choice'.
d
peeding up positive developments can be considered a promising step in improving
akers and scientists; this report has been written to
and the question whether it i
will need assistance in making safe decisions. This is done, for example, in the
Sustainable Safety vision (Wegman & Aarts, 2005) and is also the basis for the safe
system approach (OECD/ITF, 2008). This approach simply eliminates dangerous
behaviour (almost completely). It prevents overtaking on a rural road, forbids
pedestrians to walk alongside the motorway which as a result they rarely do in
Europe. But this will not always be possible and then it can be tried to catch and
punish the offender (a proven method) or to entice him or her into making safe
choices. Incentives can be used to achieve this, for example a green wave of traffic
lights if one keeps to the limit. Recently David Ward in his speech for an OECD
conference called it 'to nudge the road user'. He based his idea on the book written
by Thaler & Sunstein (2008), entitled Nudge. The cover of this book shows a mature
elephant gently pushing a baby elephant forward. The baby elephant can actually do
what he wants but is invited, be it insistently by the mature elephant's trunk, to take
those decisions that really are good for him: 'c
b
2.5. Conclusions
Benchmarking the road safety performance of countries as a basis for learning an
s
road safety. Mainly because of the simplicity of the approach and the appeal to a
wider audience, amongst which politicians and policymakers, benchmarking is
already applied in many fields, but not really in the world of road safety yet. A simple
ranking, but even better, a well-accepted benchmarking could result in inviting
experts to explain positions and to explain changes in positions to a wider audience.
Without any doubt, such benchmark results will attract attention from the media and
this can be used to make further steps. However, if we consider benchmarking
mainly as a basis for learning from each other, we are not only interested in the final
score/rate/ranking, but in the backgrounds of those scores, in the components that
contribute to the scores, and in the potential for improvements. Benchmark results
need to be accepted by policy m
obtain support for this in the field of road safety.

Benchmarking is a process in which countries or sub-national jurisdictions (states,
provinces, 'lnder', etc.) evaluate various aspects of their performance in relation to
other practices, among which the so-called 'best in class'. The benchmark results
25
enable countries or jurisdictions to learn from others as a basis for developing
measures and programmes which are aimed at increasing their own performance.
o be able to carry out meaningful benchmarking, performance indicators have to
be designed. The advantages of working with these indicators are imminent:
simplification, quantification and communication. In SUNflowerNext we distinguish
three types of indicators covering all elements of the SUNflower pyramid: road
safety performance indicators, policy performance indicators and implementation
performance indicators. These three indicators should be combined into a
composite index. A comparison of the pros and cons led to the conclusion that it
would be attractive to develop a composite index for road safety. This means that a
composite index for road safety must be developed in which all components of the
SUNflower pyramid are represented and in which attention should be paid to the
cons discussed in this chapter.
By developing a composite index for road safety (the SUNflower road safety
performance index) it became apparent that we still lack for knowledge to include all
relevant aspects. Nevertheless, it is considered feasible to study performance
indicators based on information from all EU Member States. It is recommended to
develop valid and reliable indicators for policy performance and for implementation
performance. This chapter presents information which suggests in which directions
these developments could take place. In addition it provides a basis and indicates
the directions for the chapters to follow in this study.



T
26
3. Designing a composite index for road safety
The number of indicators suggested for use in the field of road safety has been
growing rapidly, especially over the last decade (e.g. ETSC, 2001; Wegman et al.,
2005; Al Haji, 2005; Hakkert et al., 2007; Hermans et al., 2008). The purposes of
road safety indicators are to enable meaningful national or sub-national (e.g.
regional, local etc.) comparisons and monitoring through time of road safety
developments. Recognising the complex character of the road safety phenomenon,
today more and more indicators are intended to measure the factors contributing to
accidents, to identify conditions which are associated with increased accident/injury
risks, and to detail the structure of traffic injury patterns. In contrast, the traditional
approach mainly considered the safety outcomes in terms of fatalities per head of
population, vehicle fleet or exposure.

Moreover, road safety is steadily developing as a major policy area (Peden et al.,
2004), where safety performance indicators could and should serve as supportive
tools for policymakers. In comparing the safety achievements of countries there is a
need to reduce the dimensions of the problem and to be able to work with a
composite index which can describe all the relevant components in a concise and
comprehensive way.

A number of studies were recently carried out that were aimed at the development
of a composite road safety index. Al Haji (2005) suggested a Road Safety
Development Index (RSDI) and used it for a comparison of road safety progress in
ten Asian countries plus Sweden. The RSDI development was started with the
definition of eight dimensions of the road safety domain, which are traffic risk,
personal risk, vehicle safety, road situation, road user behaviour, socio-economic
background, road safety organization and enforcement. For each dimension, one or
several quantitative indicators were suggested and their applicability was analysed
based on available data. For example, the road variable was defined as the
percentage of paved roads out of the total road network; road user behaviour was
defined as the percentage of seatbelt use and helmet use, and so on. Finally, each
country was characterized by eleven separate indicators which were combined into
one composite index. To make a composite index, Al Haji (2005) applied three ways
of weighting, which were 1) the simple equal average, 2) the use of theoretical
weights, and 3) the principal component analysis. The results of the different
methods were consistent and enabled a robust classification of countries into three
groups of high, medium or low safety development.

Hermans et al. (2008) studied the issue of assigning weights to individual indicators,
to provide a combined road safety index. The researchers considered the seven
safety domains which were defined by the SafetyNet project (Hakkert et al., 2007):
alcohol and drugs, speed, protective systems, visibility (daytime running lights),
vehicles, infrastructure, and trauma care. They suggested one indicator for each
domain. Based on the data available in the international databases, from the World
Health Organisation, and from the SARTRE project, the indicators were estimated
for 21 European countries. Five weighting techniques as suggested by Nardo et al.
(2005) were used to combine the separate indicators into one index: factor analysis,
analytical hierarchy process, budget allocation, data envelopment analysis, and
equal weighting. The rankings resulting from these weighting methods were further
compared with the countries' rankings according to the personal safety (the number
of traffic fatalities per million inhabitants). It was found that different weighting
27
methods agreed most on the ranking of countries with a low road safety ranking.
Moreover, of the five methods, the data envelopment analysis method resulted in a
ranking which best approaches the countries' ranking based on the personal safety.

Both studies have clearly demonstrated the possibilities for creating composite road
safety indices. However, they both considered a relatively small number of basic
indicators, for some of which the quality of the data for the quantitative measures
selected can be questioned. At the same time, the limitations of basic indicators
used in these analyses were probably caused by the lack of real data on other,
theoretically more suitable indicators.

In the current project, we aim to create and explore a comprehensive composite
road safety index, based on the recent concepts of the road safety domain
developed by the SUNflower (Wegman et al., 2005) and SafetyNet (Hakkert et al.,
2007) projects. The types of benchmarking discussed in the literature will be
considered, and data on the European countries that were collected in the SafetyNet
project and which are available from international databases, will be used.

The SUNflower approach described the road safety domain as a pyramid consisting
of several layers, from bottom to top: safety measures and programmes (as the road
safety policy performance); safety performance indicators (as intermediate
outcomes); the numbers of accident fatalities/injuries (as the final outcomes) and the
social costs of accidents/injuries at the very top. An additional 'Structure and culture'
layer has been added at the bottom of the pyramid to include the background
conditions of the system or the policy context (Koornstra et al., 2002).

The reason for the development of safety performance indicators (SPIs) is the
assumption that accidents and injuries are only the tip of the iceberg, because they
occur as the 'worst case' result of unsafe operational conditions in the road traffic
system. At the same time, those who are responsible for road safety need to take
into account as many factors influencing safety as possible or, at least, those factors
that they are able to affect or control. Hence, additional safety performance
indicators (besides accident/ injury numbers) are required to provide a means for
monitoring the effectiveness of the safety actions that are taken. Safety performance
indicators can be seen as measures that are causally related to accidents or injuries
and are used in addition to the figures about accidents or injuries, in order to
indicate safety performance or understand the processes that lead to accidents
(ETSC, 2001).

The SafetyNet project (Hakkert et al., 2007) provided a further methodological basis
for the SPIs' development. A precondition in the development of SPIs was that they
should be able to reflect unsafe operational conditions of the road traffic system and
should, therefore, be more general than the direct outputs of specific safety
interventions. Based on the potential of different road safety domains for increasing
road safety as well as on the experiences and data available, seven problem areas
were designated as central to road safety activities in Europe and were selected for
the development of SPIs. They are:

Alcohol and drug-use
Speeds
Protective systems
Daytime running lights (DRL)
Vehicles (passive safety)
28
Roads
Trauma management

According to Hakkert et al. (2007), SPIs that are developed for a certain safety
domain should reflect the factors contributing to road accidents/injuries and
characterize the scope of the problem identified. The development of SPIs begins
with a definition of the problem (i.e. the operational conditions of the road traffic
system which are unsafe and result in accidents/fatalities as the 'worst case') and
continues with the conversion of this information into measurable variables. Using
the data provided by the national representatives of the 27 EU Member States plus
Norway and Switzerland, SPIs were developed and country comparisons were
performed see Vis & Van Gent (2007).
3.1. Basic indicators
This study sets out to develop a composite road safety performance index for
benchmarking purposes, which combines all layers of the road safety pyramid. In
Section 2.2 three types of indicators were defined:
1. Road safety performance indicators (quality of road safety);
2. Policy performance indicators (quality of road safety policies);
3. Implementation performance indicators (quality of implementation of road safety
policy).

We still have to develop sound Implementation performance indicators, in which we
would like to use the distinction as indicated by Bliss & Breen (to be published):
institutional management functions and interventions. This was described in Chapter
2, and we recommend to develop these indicators and to collect the necessary data.

For Policy performance indicators we developed a set of five indicators (A1-A5). In
Road safety performance indicators we made a distinction between final outcome
(B1-B7) and intermediate outcome (C1-C7).

A fourth group of indicators (D) was added trying to present some background
variables for each country, as a first attempt to identify components of the lowest
level of the pyramid: Structure and culture (D1-D2).

To develop a composite index, basic indicators should first be defined, for each
layer of the pyramid that needs to be considered.

In the present context, the benchmarking of road safety policies (the A-group of
indicators) is described in terms of the quality of national road safety plans. Five
components are included and analysed:

A1 Safety targets the availability and ambition of quantitative national safety
targets;
A2 Selection of interventions whether a sound analysis preceded the
development of the national safety programme;
A3 Economic evaluation whether a sound economic evaluation preceded the
design of the national safety programme;
A4 Monitoring whether the national safety programme is systematically
monitored;
A5 Stakeholders who is responsible for the programme's performance.
29
The values (categories) for each indicator are presented in Table 3.1. When a
national road safety policy is characterized (safety measures and programmes in the
pyramid), for each indicator one of the categories is selected. The values selected
for each indicator (a, b, c or d) are based on the national road safety programme,
other background papers and available follow-up reports, including those prepared
by the ETSC, OECD or other international working groups.

Indicators Possible values
A1 Safety targets a. Ambitious
b. Available but not ambitious
c. Not available
A2 Selection of interventions a. Sound analysis and diagnosis of road safety problems
preceded the programme's development, and evidence-based
interventions were selected
b. Some analysis was performed, and evidence-based
interventions were selected
c. Detailed analysis of road safety problems was performed,
however, the selection of interventions was arbitrary
d. The diagnosis of road safety problems was poor and the
selection of interventions was arbitrary
A3 Economic evaluation a. Sound economic evaluation preceded the programme's
composition
b. Some economic evaluation was performed
c. Economic evaluation was not performed
A4 Monitoring the
programme's performance
a. Systematic monitoring takes place
b. A need for monitoring is stated but monitoring reports are not
found
c. No evidence of monitoring activities
A5 Programme's stakeholders a. Commitment was stated on the governmental level, and the
programme is supervised by a central authority which is
empowered to coordinate the activities of all other bodies
b. No commitment from the government, however, a central
authority was commissioned for the programme's performance
c. A number of authorities share the responsibility for the
programme's performance
d. No authority has a clear responsibility for the programme's
performance
Table 3.1. Definition of basic indicators for the A-group: characteristics of national
safety policies (Policy performance indicators).


The benchmarking of road safety performance, the B-group of indicators, deals with
the final outcomes of the system, i.e. the numbers of road crash fatalities and
injuries, which should be presented in a form suitable for comparisons. The present
study focuses on four issues: personal safety, traffic safety (rates and risks), the
scope of traffic injury, and the scope of the problem of vulnerable road users. For
each issue one or several indicators are defined, see Table 3.2.

30
Issues considered Indicators defined*
Personal safety B1 Fatalities per million inhabitants
Traffic safety rate
Traffic safety risk
B2 Fatalities per million passenger cars
B3 Fatalities per 10 billion passenger-km travelled
Scope of traffic injury B4 Injury accidents per fatality**
Scope of the problem of
vulnerable road users
B5 Share of pedestrian fatalities out of the total fatalities
B6 Share of bicyclist fatalities out of the total fatalities
B7 Share of motorcyclist fatalities out of the total fatalities
*Note: Obviously, various definitions are possible. B2 can for instance be defined per total vehicle fleet
instead of passenger cars, B3 per vehicle kilometre travelled instead of passenger kilometre, etc. The
selection was based on the estimates available to the EU Member States.
** Considering the reporting problems and definition's differences among the countries, a preferable
indicator would be the 'number of hospitalized injuries per 1 fatality', which is unavailable as yet.
Table 3.2. Definition of basic indicators for the B-group: Road safety performance
indicator, final outcomes.


The second group of Road safety performance indicators, the C-group of indicators,
captures the intermediate outcomes, containing the safety performance indicators
which characterize the safety quality of the road traffic system. Having analysed the
data availability in the seven pre-defined SPI areas (Vis & Van Gent, 2007) as well
as the summaries on safety performance indicators (PIN Flashes) recently
published by the ETSC, it was decided to limit the present project to only those
safety areas for which the estimates are available for a significant number of the
Member States. The three remaining safety areas are: alcohol-impaired driving, use
of protective systems in cars, and vehicles (passive safety) which includes the
crashworthiness of the passenger car fleet and the vehicle fleet composition. For
each area one or two indicators have been defined, see Table 3.3.

Safety areas considered Indicators defined
Alcohol-impaired driving C1 Share of total for fatalities in drink-driving accidents
Use of protective systems
in cars
C2 Daytime wearing rates of seatbelts in the front seats (aggregated
for driver and front passenger)
C3 Daytime wearing rates of seatbelts in the rear seats
Vehicles:
Crashworthiness of the
passenger car fleet

Vehicle fleet composition

C4 Average EuroNCAP score of passenger car fleet
C5 Median age of the passenger car fleet

C6 Share of motorcycles in the vehicle fleet
C7 Share of heavy goods vehicles (HGV) in the vehicle fleet
Table 3.3. Definition of basic indicators for the C-group: Road safety performance
indicator, intermediate outcomes, SPIs.

In addition, two background indicators were added (D-group) to characterize the
motorization level and the population density of the country, see Table 3.4.

Characteristic considered Indicators defined
Motorization level D1 Number of passenger cars per 1000 inhabitants
Population density D2 Population per 1 km
2
of country's territory
Table 3.4. Definition of basic indicators for D-group: background characteristics.
31
32
The choice of indicators and their definitions have a preliminary character. The
choice of indicators used was also influenced by the immediate availability of data.
The aim of this chapter is to demonstrate along which lines a composite index can
be designed.
3.2. Data collection
In total, 21 basic indicators were defined for consideration. The estimates of the
indicators and/or data for their calculation were taken from a wide range of
international databases and recent publications of international working groups,
including:

For group A OECD/ITF (2008), OECD/ECMT (2008);

For group B EC (2007), ERSO (2008);

For C1 ETSC (2007a); for C2, C3 - ETSC (2007b); for C4 - Vis & Van Gent
(2007); for C5 - UNECE (2008); for C6, C7 - EC (2007) and OECD country reports
(2008).

Table 3.5 contains the basic indicators which were estimated for 27 countries: 25
Member States of the European Union plus Norway and Switzerland. The majority
of indicators are not yet available for the new Member States Bulgaria and Romania.
3.3. Method of analysis
The purpose of this analysis is to create a composite road safety performance index
and to explore the similarities in basic indicators. The composite index will enable
ranking countries in accordance with their safety performance or, at least, to define
several groups of countries with different levels of safety performance.

As can be seen in Table 3.5, only six countries have values for the whole set of
basic indicators A to D. Therefore, prior to the analysis, missing data imputations
should be performed. A description of the imputation method and the final data set,
with initial and imputed values of basic indicators, are presented in Appendix 1.

The initial examination revealed that the effect of the Malta data was very strong and
that it should be considered an outlier. Hence, Malta was excluded from the com-
posite index' building process. However, using the factors developed, the scores -
combined indices - for Malta can still be calculated, and, consequently, these appear
in the final plots.



C
o
u
n
t
r
y

C
o
d
e

B
1
B
2
B
3

B
4

B
5
B
6
B
7
D
1
D
2
C
1
C
2

C
3
C
4
C
5
C
6
C
7
A
1
A
2
A
3

A
4

A
5





F
a
t
a
l
i
t
i
e
s

p
e
r

m
l
n

i
n
h
a
b
i
-
t
a
n
t
s
-
2
0
0
6

F
a
t
a
l
i
t
i
e
s

p
e
r

m
i
l
l
i
o
n

p
a
s
s
e
n
g
e
r

c
a
r
s
-
2
0
0
6

F
a
t
a
l
i
t
i
e
s

p
e
r

1
0

b
i
l
l
i
o
n

p
k
m
-
2
0
0
6

I
n
j
u
r
y

A
c
c
i
d
e
n
t
s

p
e
r

f
a
t
a
l
i
t
i
e
s
-
2
0
0
6

P
e
d
e
s
t
r
i
a
n

f
a
t
a
l
i
t
i
e
s

a
s

%

o
f

t
o
t
a
l
-
2
0
0
6

B
i
c
y
c
l
i
s
t

f
a
t
a
l
i
t
i
e
s

a
s

%

o
f

t
o
t
a
l
-
2
0
0
6

M
o
t
o
r
c
y
c
l
i
s
t

+

m
o
p
e
d

f
a
t
a
l
i
t
i
e
s

a
s

%

o
f

t
o
t
a
l
-
2
0
0
6

N
u
m
b
e
r

o
f

p
a
s
s
e
n
g
e
r

c
a
r
s

p
e
r

1
0
0
0

i
n
h
a
b
i
t
a
n
t
s
,

2
0
0
6

P
o
p
u
l
a
t
i
o
n

p
e
r

1

k
m
2

o
f

a
r
e
a
-
2
0
0
6

S
h
a
r
e

o
f

d
r
i
n
k
-
d
r
i
v
i
n
g

a
c
c
i
d
e
n
t
s
-
2
0
0
5

W
e
a
r
i
n
g

r
a
t
e
-
F
r
o
n
t

s
e
a
t
s

a
g
g
r
e
g
a
t
e
d
-
2
0
0
5

W
e
a
r
i
n
g

r
a
t
e
-
R
e
a
r

s
e
a
t
s
-
2
0
0
5


A
v
e
r
a
g
e

E
u
r
o
N
C
A
P

s
c
o
r
e
-
2
0
0
3

M
e
d
i
a
n

a
g
e

o
f

p
a
s
s
e
n
g
e
r

c
a
r
s
-
2
0
0
3

%

M
o
t
o
r
c
y
c
l
e
s

i
n

f
l
e
e
t
-
2
0
0
6

%

H
G
V

i
n

f
l
e
e
t
-
2
0
0
6

S
a
f
e
t
y

t
a
r
g
e
t
s

S
e
l
e
c
t
i
o
n

o
f

i
n
t
e
r
v
e
n
t
i
o
n
s

E
c
o
n
o
m
i
c

e
v
a
l
u
a
t
i
o
n

M
o
n
i
t
o
r
i
n
g

S
t
a
k
e
-
h
o
l
d
e
r
s

A
u
s
t
r
i
a

A
T

8
8
1
7
5
1
0
0

5
4
.
6

1
5
.
1
%
6
.
6
%
1
8
.
4
%
5
0
7

9
8
.
9
6
.
0
%
8
3

5
2
n
/
a
7
.
5
1
2
.
4
%
7
.
0
%
a
c
b

b

a

B
e
l
g
i
u
m

B
E

1
0
1
2
1
6
9
6

3
8
.
5

1
1
.
4
%
8
.
6
%
1
5
.
5
%
4
7
0

3
4
7
.
0
4
.
4
%
7
1

n
/
a
2
1
.
3
6
.
7
6
.
0
%
1
1
.
1
%
b
b
b

b

b

S
w
i
t
z
e
r
l
a
n
d

C
H

5
0
9
5

n
/
a

5
8
.
1

2
0
.
5
%
9
.
5
%
2
1
.
6
%
5
1
9

1
8
1
.
8
1
9
.
3
%
8
2

5
3
n
/
a
7
.
3
1
2
.
5
%
6
.
4
%
a
a
a

b

a

C
y
p
r
u
s

C
Y

1
1
1
2
3
6
1
6
7

3
1
.
1

2
2
.
1
%
2
.
3
%
2
9
.
1
%
4
7
9

8
3
.
8
2
2
.
5
%
8
0

n
/
a
n
/
a
9
.
3
7
.
6
%
2
1
.
7
%
n
/
a
n
/
a
n
/
a

n
/
a

n
/
a

C
z
e
c
h

R
e
p
u
b
l
i
c


C
Z

1
0
4
2
6
4
1
4
6

2
0
.
8

1
9
.
0
%
1
0
.
3
%
1
0
.
9
%
3
9
9

1
3
0
.
4
5
.
5
%
7
2

1
3
2
0
.
7
1
1
1
5
.
1
%
9
.
0
%
b
c
b

b

b

G
e
r
m
a
n
y

D
E

6
2
1
1
0
5
7

6
4
.
4

1
4
.
0
%
9
.
5
%
1
7
.
7
%
5
6
6

2
3
0
.
6
5
.
1
%
9
6

8
9
2
0
.
1
6
1
0
.
5
%
5
.
1
%
a
b
b

a

b

D
e
n
m
a
r
k

D
K

5
6
1
5
4
5
6

1
7
.
7

1
9
.
6
%
1
0
.
1
%
1
4
.
7
%
3
7
1

1
2
6
.
4
2
3
.
0
%
8
5

6
3
n
/
a
7
.
3
6
.
7
%
1
8
.
7
%
a
b
b

b

b

E
s
t
o
n
i
a

E
E

1
5
2
3
8
9
1
7
5

1
2
.
7

2
7
.
2
%
4
.
1
%
4
.
1
%
4
1
3

2
9
.
7
2
8
.
4
%
7
4

3
0
n
/
a
1
2
.
7
1
.
9
%
1
4
.
0
%
b
c
b

b

a

G
r
e
e
c
e

E
L

1
4
9
3
7
5
1
7
4

9
.
8

1
6
.
1
%
1
.
3
%
3
0
.
0
%
4
0
7

8
4
.
6
1
0
.
7
%
n
/
a

n
/
a
2
0
.
7
8
1
7
.
2
%
1
7
.
4
%
a
b
b

b

b

S
p
a
i
n

E
S

9
3
2
0
1
1
1
7

2
4
.
3

1
5
.
3
%
1
.
8
%
1
7
.
6
%
4
6
4

8
7
.
9
n
/
a
7
4

5
1
n
/
a
7
.
6
1
4
.
6
%
1
6
.
7
%
a
c
b

b

b

F
i
n
l
a
n
d

F
I

6
4
1
3
6
5
3

2
0
.
1

1
1
.
9
%
1
1
.
3
%
9
.
5
%
4
7
5

1
5
.
6
2
3
.
5
%
8
8

7
8
n
/
a
9
.
1
1
0
.
5
%
1
1
.
6
%
a
b
b

b

b

F
r
a
n
c
e

F
R

7
7
1
5
3
6
4

1
7
.
1

1
1
.
4
%
3
.
8
%
2
3
.
5
%
5
0
4

1
1
3
.
1
2
8
.
8
%
9
7

7
0
n
/
a
8
.
3
6
.
4
%
1
3
.
7
%
a
a
b

a

a

H
u
n
g
a
r
y

H
U

1
2
9
4
4
6
2
7
1

1
6
.
1

2
2
.
7
%
1
1
.
7
%
1
0
.
1
%
2
9
3

1
0
8
.
2
8
.
8
%
6
7

3
4
2
0
.
6
8
3
.
7
%
1
2
.
5
%
a
c
b

b

b

I
r
e
l
a
n
d

I
E

8
6
2
1
1
1
3
0

1
8
.
6

1
9
.
0
%
3
.
0
%
1
6
.
3
%
4
1
8

6
1
.
4
n
/
a
8
6

4
6
n
/
a
5
.
1
1
.
6
%
1
4
.
7
%
b
b
b

a

b

I
t
a
l
y

I
T

9
6
1
6
2
7
4

4
2
.
0

1
2
.
6
%
4
.
8
%
2
3
.
8
%
5
9
7

1
9
6
.
3
n
/
a
7
1

n
/
a
n
/
a
7
.
9
2
0
.
5
%
8
.
7
%
c
d
c

c

d

L
i
t
h
u
a
n
i
a

L
T

2
2
4
4
9
8
1
9
1

8
.
9

3
4
.
0
%
n
/
a
n
/
a
4
7
0

5
1
.
9
1
1
.
8
%
6
0

n
/
a
n
/
a
1
4
.
5
1
.
4
%
7
.
7
%
b
c
b

b

b

L
u
x
e
m
b
o
u
r
g

L
U

7
6
1
1
6
5
4

2
1
.
2

9
.
7
%
1
.
6
%
0
.
0
%
6
6
1

1
8
3
.
1
0
.
0
%
8
0

6
0
n
/
a
4
.
4
1
0
.
0
%
8
.
0
%
n
/
a
n
/
a
n
/
a

n
/
a

n
/
a

L
a
t
v
i
a


L
V

1
7
8
5
2
0
2
5
6

1
0
.
6

3
8
.
0
%
6
.
0
%
4
.
0
%
3
6
0

3
5
.
3
2
1
.
7
%
7
7

n
/
a
2
1
.
3
1
0
3
.
7
%
1
2
.
2
%
b
c
b

b

b

M
a
l
t
a

M
T

2
5
4
6
4
9

8
4
.
1

3
5
.
3
%
0
.
0
%
1
7
.
6
%
5
3
5

1
3
6
0
.
0
n
/
a
9
6

2
8
1
8
.
6
9
4
.
4
%
1
6
.
4
%
a
c
b

b

b

N
e
t
h
e
r
l
a
n
d
s

N
L

4
5
1
0
2
4
8

3
3
.
6

9
.
4
%
1
8
.
3
%
1
8
.
4
%
4
4
2

3
9
4
.
2
1
5
.
3
%
9
0

6
4
2
0
.
2
6
6
.
5
%
1
1
.
3
%
a
a
b

a

a

N
o
r
w
a
y

N
O

5
2
1
1
6
n
/
a

3
2
.
7

1
4
.
5
%
3
.
3
%
1
5
.
3
%
4
4
5

1
4
.
4
n
/
a
9
1

8
3
2
1
8
9
.
4
%
1
7
.
0
%
b
a
a

b

b

P
o
l
a
n
d


P
L

1
3
7
4
0
8
2
3
5

8
.
9

3
2
.
3
%
1
1
.
1
%
3
.
9
%
3
5
1

1
2
1
.
9
9
.
8
%
7
8

n
/
a
n
/
a
1
1
.
8
4
.
7
%
1
4
.
4
%
b
c
b

b

b

P
o
r
t
u
g
a
l

P
T

9
2
2
2
8
1
3
1

3
6
.
8

1
6
.
0
%
4
.
1
%
2
4
.
1
%
4
0
5

1
1
5
.
3
n
/
a
8
6

4
5
n
/
a
9
.
7
9
.
0
%
2
1
.
3
%
b
c
b

b

b

S
w
e
d
e
n

S
E

4
9
1
0
7
4
5

4
0
.
7

1
2
.
4
%
5
.
8
%
1
5
.
7
%
4
6
1

2
0
.
3
3
4
.
0
%
9
2

7
3
2
2
.
6
7
9
.
6
%
9
.
2
%
a
a
a

a

a

S
l
o
v
e
n
i
a

S
I

1
3
1
2
7
0
1
1
3

4
4
.
4

1
3
.
7
%
5
.
3
%
2
0
.
5
%
4
8
8

9
9
.
0
3
2
.
2
%
8
7

3
0
n
/
a
5
.
0
4
.
8
%
6
.
3
%
b
c
b

b

b

S
l
o
v
a
k
i
a

S
K

1
0
7
4
3
9
2
1
5

1
3
.
8

9
.
3
%
5
.
5
%
7
.
4
%
2
4
7

1
1
0
.
5
1
2
.
0
%
n
/
a

n
/
a
n
/
a
n
/
a
3
.
7
%
1
1
.
9
%
b
c
b

b

b

G
r
e
a
t

B
r
i
t
a
i
n

U
K

5
4
1
1
6
4
8

5
8
.
9

2
1
.
0
%
4
.
6
%
1
7
.
5
%
4
7
1

2
4
9
.
3
1
7
.
5
%
9
0

8
4
2
0
.
6
5
3
.
7
%
1
0
.
7
%
a
b
b

a

a

T
a
b
l
e

3
.
5
.

B
a
s
i
c

i
n
d
i
c
a
t
o
r
s

e
s
t
i
m
a
t
e
d

f
o
r

2
7

E
u
r
o
p
e
a
n

c
o
u
n
t
r
i
e
s

(
n
/
a

=

n
o
t

a
v
a
i
l
a
b
l
e
)
.

To combine the basic indicators into a composite one, weights based on statistical
models are applied. We examined both Principal Component Analysis (PCA) and
Common Factor Analysis (FA) weighting (Tabachnick & Fidell, 2006). Both methods
group collinear indices to form a composite index that captures as much common
information among sub-indicators as possible. The idea under PCA/FA is to account
for the highest possible variation in the set of indicators using the smallest possible
number of factors (Nardo et al., 2005). If there is no correlation between indicators
these methods can not be used to obtain the weights.

The first step in the FA is to check the correlation structure of the data, and the
second step is to identify a certain number of latent factors, smaller than the number
of indicators, representing the data. In the PCA, we retain those factors that account
for the largest amount of variance.

We used Kaiser-Meyer-Olkin (KMO) statistics to predict and demonstrate if the data
are likely to factor well, based on their correlations and partial correlations. There is
a KMO statistic for each individual variable, and an overall KMO statistic. It is
customary that the KMO overall should be 0.60 or higher, to proceed with the factor
analysis. If this condition is not satisfied, one should drop the indicator variables with
the lowest individual KMO statistic values, until the KMO overall rises above 0.60.

In both methods (PCA/FA), the combination of factors into a composite index is
done by their weighted sum, where the weights are taken in accordance with the
variance explained by each factor. At this point, different estimation procedures are
possible. For both PCA and FA, we used orthogonal rotation
1
of factors and, hence,
'the variance explained by each factor' was applied for estimating the weights.

The Principal Component Analysis and the Common Factor Analysis were used
because both PCA and FA are suitable procedures for analysis. Since it is not clear
beforehand which one is the most suitable, five trials of creating a composite index
were performed. The reasons why we decided to make five trials are explained in
more detail later in this chapter. The five types of analysis are:

1) PCA-all a PCA in which all the variables (basic indicators) were
analysed together.
2) PCA-groups a PCA in which each group of basic indicators (A, B, C, D)
was first analysed separately to create the group factors.
The group factors were then analysed together to provide
the final composite index.
3) FA-4Factors a common factor analysis with all variables considered and
four factors' solution accepted.
4) FA-2Factors-noC4 a common factor analysis with two factors' solution, where
C4 (average EuroNCAP score) is excluded from the
analysis, due to statistical reasons.
5) FA-2factors a common factor analysis with two factors' solution, where
C4 is retained in the analysis.

Note: in trials 3-5, variables B6 (share of bicyclist fatalities), B7 (share of
motorcyclist fatalities), C7 (percentage of HGV in the vehicle fleet) were excluded
from creating a combined indicator due to statistical reasons (low values of the KMO
statistics).

1
VARIMAX: change of coordinates in PCA that maximize the sum of the variance of the loading
vectors
34
Once the factors are created, the following procedure is applied to calculate a
composite index, for each country:
a. Standardize the data, for each country, i.e. subtract mean and divide by standard
deviation, for each variable.
b. Multiply these standardized variables by their respective standardized scoring
coefficients.
c. Sum up these products over all the relevant variables. The sum is the value of
the new variable Factor 1.
d. Repeat steps a-c for Factor 2, Factor 3, etc (if relevant). This creates the data set
scores.
e. The composite index is a weighted sum of the factors.

The results of each of the five trials made it possible to produce a combined safety
index (called WF weighted factor or weighted index) for each country, as well as
for the clusters of countries with similar values of the combined index. In addition,
the development of a composite index provided an insight into similarities and
dissimilarities in the behaviour of basic indicators (when the indicators' involvement
in building the factors is considered).

Detailed results of each of the five analyses are presented in Appendix 2. In
addition, Appendices 3 and 4 present the tools that were accordingly produced by
the PCA and FA analyses for the estimation of country scores. Section 3.4 gives a
brief summary of the findings. Comparisons of the results of the five trials, including
considerations of their meanings and practical implications for the development of a
composite road safety performance index, are further discussed in Section 3.5.
3.4. Summary of findings
As mentioned earlier, the statistical methods applied (PCA/FA) group together basic
indicators that are collinear to form a composite index, which should capture as
much of common information among basic indicators as possible. The idea under
PCA/FA is to account for the highest possible variation in the set of indicators using
the smallest possible number of factors. As a result of the analysis, a number of
factors are fitted to the data, where each factor presents a composition of basic
indicators. Considering the factors created one can see which variables (basic
indicators) contribute more to each one of the factors. Moreover, considering the
'safety-desirable' behaviour of basic indicators and their coefficients when the factor
values are estimated, one can state whether higher or lower values of each factor
are associated with better safety performance (see Appendix 2).

1) PCA-all

In the PCA-all analysis, five factors were fitted to the data. Factor 1 mainly reflects
the road safety performance indicators, car fleet's age and seatbelt use. Factor 2
mainly reflects the policy performance indicators but also includes a negative
correlation with C1 (share of drink-driving accidents). Factor 3 reflects the share of
bicyclist fatalities, EuroNCAP scores for cars and population density. Factor 4
reflects the share of motorcycles in the fleet and the share of motorcyclist fatalities.
And Factor 5 reflects the share of HGVs in the fleet, the number of injury accidents
per fatality and the motorization level of a country.

35
When separate indicators are plotted against the factors, similarities in their
behaviour can be recognized. This way, similar behaviour was noted for variable
groups such as: B1-B2-B3 (number of fatalities per head of population, vehicles,
passenger kilometres travelled), A1-A4 (safety targets, monitoring the programme
performance), A2-A3-A5 (other characteristics of safety programmes) and C2-C3
(safety belt wearing rates in front and rear seats).

Furthermore, the factors built are weighted together to provide a composite index
(weighted index WF). The weighted index can be applied to countries' ranking.
Furthermore, using the WF and a WARD clustering procedure, a classification tree
can be produced, i.e. the countries can be classified into similar groups. Inside the
group the WF values are close, but there are distances between the groups (The
classification trees produced by the WARD procedure are presented in Appendix 2).
Using a classification tree, various country groups (clusters) can be defined,
depending on the level of 'distances' between the countries in the same group,
which is selected as a threshold.

For example, Figure 3.1 presents the countries' subdivision into six clusters based
on the WF values. The country positions in Figure 3.1 are plotted using the WF
values and Factor 1 values. According to this classification, the countries with the
best safety level are Sweden, Great Britain and France; the second group includes
Luxembourg, Norway, Germany and Ireland; the third group consists of Slovenia,
Finland and the Netherlands; the fourth group of Switzerland, Austria, Denmark and
Estonia; the fifth group of Cyprus, Spain, Belgium, Slovakia, Portugal, Malta and
Latvia; and the sixth group consists of Lithuania, Greece, Hungary, Poland, Czech
Republic and Italy.

PCA - all variables together
Countries-WF*Factor1(out of 5), using weighted factor for clustering
-2
-1 -0,5 0 0,5
WF
-1
0
1
F
a
c
t
o
r

1
2
3
Sweden Norway
Slovenia
Finland
Switzerland
1 2 3 4 5 6 Cluster membership
Great Britain
France
Ireland
Germany
Netherlands
Luxembourg
Denmark
Estonia
Austria
Cyprus
Belgium
Slovakia
Czech
Republic
Hungary
Greece
Malta
Latvia
Portugal Italy
Lithuania
Spain
Poland

Figure 3.1. Countries plotted using the WF values and Factor 1 values (PCA-all
analysis).



36
2) PCA-groups

The first step in the PCA-groups analysis was to fit one combined factor for each
group of indicators separately, namely for the A, B, C and D groups of basic
indicators. The four combined factors obtained are then subjected to another PCA,
where the final composite index is generated.
the C-group indicators (road safety performance intermediate outcome),
w communalities with other indicators were observed for C6, C7 indicators (the
percentages of motorcycles and HGV in the vehicle fleet, respectively), which,
consequently, were excluded from the analysis. For the retained indicators, two
factors were chosen, where variables C2-C3 (seatbelt wearing rates) and C5
(median age of passenger cars) contributed mostly to Factor 1, while C1 (share of
drink-driving accidents) and C4 (average EuroNCAP score) contributed mostly to
Factor 2.

For the D-group indicators (background characteristics), one factor was chosen,
which reflected both motorization level and population density.

In the second step of this analysis two factors were chosen, where the combined
factors of A- and C-groups composed Factor 1, and the combined factors of B- and
D-groups composed Factor 2. Plotting the positions of the initial group factors and
the positions of the countries on the dimensions of these two combined factors
clearly showed a lack of similarity in the behaviour of the group factors. For the
countries one general 'cloud' was noted, with several outsiders (e.g. Italy, Malta),
see Figure 3.2. Inside the 'cloud', the countries with better safety-related positions
are: France, Great Britain, Sweden, Switzerland, the Netherlands, Germany and
Norway.


For the A-group indicators (policy performance), one factor was fitted to all basic
(A1-A5) indicators in the first step of this analysis.

For the B-group indicators (road safety performance final outcome), two factors
were chosen where variables B1-B2-B3-B4-B5 (number of fatalities per head of
population, vehicles, passenger kilometres travelled, injury accidents per fatality,
and share of pedestrian fatalities) contribute to Factor 1 and B6-B7 (shares of
bicyclist and motorcyclist fatalities, accordingly) contribute to Factor 2. The results
demonstrated similarities in the behaviour of B1-B2-B3-B5 indicators and very
different behaviours of B4, B6, B7.

Exploring
lo
37

PCA-group ABCD (2 factors chosen)
Countries-Factor2*Factor1only
5
7,5
MT
-2,5
0
2,5
F
a
c
t
o
r

2
SE
LT
PL
HU
SK
IE
CY
SI
DK
FI
UK
CH
NL
DE
LV
AT
CZ
EE
EL
PT
ES
BE
LU
IT
NO
FR
-3 -2 -1 0 1 2 3
Factor 1

Figure 3.2. Countries' positions on the dimensions of the combined factors
ombining FA, FB, FC, and FD).

(c

The three trials of creating a composite index using FA considered 18 variables
asic indicators), where B6 (share (b of bicyclist fatalities), B7 (share of motorcyclist
talities) and C7 (percentage of HGV in fleet) were excluded due to their low values
rogrammes) and C1 (share of drink-driving accidents).

consists of Slovakia, Italy, Poland,
fa
in the KMO statistics.

3) FA with four factors

In the FA with four factors the basic indicators B1-B3 (number of fatalities per head
of population, vehicles, passenger kilometres travelled), B5 (share of pedestrian
fatalities), C5 (median age of passenger cars), C2-C3 (safety belt wearing rates in
front and rear seats), B4 (injury accidents per fatality), and D1 (number of passenger
cars per head of population) provided a major contribution to Factor 1, whereas
Factor 2 reflected mostly the behaviour of policy performance indicators A1, A2, A4,
A5 and C6 (percentage of motorcycles in vehicle fleet). Factor 3 consisted of the C4
variable only (average EuroNCAP score), whereas Factor 4 reflected mostly the
behaviour of D2 (population density), A3 (quality of economic basis of safety
p

Based on the combined indicator (WF values) and Factor 1 values obtained in this
analysis, the countries were subdivided into five groups (Figure 3.3). The countries
with the best safety level (first group) are Sweden, Norway, Switzerland, Great
Britain, France, the Netherlands and Germany; the second group includes Finland,
Denmark, Malta, Austria, Luxembourg and Ireland; the third group consists of
Cyprus, Slovenia, Spain, Portugal and Belgium; the fourth group of Greece, Czech
epublic and Estonia; and the fifth group R
Hungary, Latvia and Lithuania.

38
Malta, Luxembourg and, especially, Italy, are 'outsiders' of their groups, having
much lower values on Factor 1 in comparison with the countries with a similar level
of the WF values.

FA - all variables together
Countries-WF*Factor1(out of 4), using weighted factor for clustering
2
3
Latvia
Hungary
Slovakia
-2
-1
1
-1 -0,5 0 0,5 1
WF
1 2 3 4 5
Cluster membership
Estonia
Poland
Czech Republic
Greece
Belgium
Portugal
Italy
Malta
Spain
Switzerland
0
F
a
c
t
o
r

1
Lithuania
Cyprus
Austria
Slovenia
Netherlands
Finland
Germany
Norway
Sweden
France
Luxembourg
Great Britain
Denmark
Ireland

Figure 3.3. Countries plotted using the WF values and Factor 1 values (FA with four
factors' solution).


4-5) FA with two factors

In the FA with two factors two alternatives were considered: including and excluding
the C4 indicator (average EuroNCAP score). This is because within the FA with four
factors, C4 variable's behaviour was nearly identical to one of the factors, where in
such cases (an additional factor consisting of only one variable) it is customary to
exclude such a variable from the analysis. The exclusion of C4 enabled fitting two
factors to the rest of the variables (indicators), but with a significantly lower value of

ping C4 inside the set and limiting the results to two
factors fitted yielded a comparable value of the explained variance (about 69%).
Therefore, we decided to present and compare the results of both trials: including

A5 (policy performance), C2-C3 (use of seatbelts) and D1-D2 (background charac-
the explained variance than in the four factors' analysis (71% versus 84%). A further
consideration revealed that kee
and excluding the C4 indicator.
In both trials of the 'FA with two factors' (with C4, average EuroNCAP score,
excluded or included), Factor 1 reflects mostly the behaviour of B1-B5 indicators
('safety product'), D1-D2 ('background characteristics'), C5 (median age of
passenger cars), C6 (percentage of motorcycles in vehicle fleet) and C4 (if included
in the analysis); whereas Factor 2 reflects mostly the behaviour of policy
performance indicators A1-A5, C2-C3 (safety belt use) and C1 (share of drink-
driving accidents) indicators.

In both cases, similar behaviour was observed for basic indicators: B1-B2-B3-B5-C5
(road safety performance final outcome - and median age of cars), A1-A2-A3-A4-
39
teristics). Moreover, C4 indicator is relatively close to the group of 'safety outcomes'
plus 'median age of cars', whereas C1 (share of drink-driving accidents) behaves
ifferent from all the other indicators. Besides, B4 (injury accidents per fatality) and
In both cases of the 'FA with two factors', the groups of countries with similar values
lows (see Figure 3.4 with the results
including C4). The countries with the best safety level (first group) are Malta, the
prus and Slovenia; the fourth
;

be
d
C6 (percentage of motorcycles in fleet) are relatively close to the D1-D2-group.

of combined indicators (WF values) were as fol
Netherlands, Sweden, Great Britain, Germany, Switzerland, France and Norway; the
second group includes Finland, Denmark, Austria, Luxembourg and Ireland; the
third group consists of Spain, Belgium, Portugal, Cy
group of Greece, Czech Republic, Italy, Slovakia, Hungary, Estonia, and Poland
and the fifth group consists of Latvia and Lithuania.
Similar to the results of the 'FA with four factors', Malta, Italy and Luxembourg
have as 'outsiders' of their groups.

FA - all variables together
Countries-WF*Factor1(out of 2), using weighted factor for clustering
1
2
-3
-2
-1
-1,5 -1 -0,5 0 0,5 1 1,5
WF
0
o
r

1
F
a
c
t
1 2 3 4 5
ech Republic
Cyprus
Slovakia
Greece
Belgium
Lithuania
Latvia
Estonia
Poland
Hungary
Cz
Portugal
Spain
Slovenia
Denmark
Italy
Malta
Austria
Ireland
Finland
Switzerland
Great Britain
Cluster membership
Norway
France
Sweden
Luxembourg
Netherlands
Germany

Figure 3.4. Countries plotted using the WF values and Factor 1 values (FA with two
factors' solution, C4 included).
3.5. Discussion and conclusions
In this chapter, we examined Principal Component Analysis (PCA) and Common
ba
iv duced
composite road safety indicator for each country, and clusters/groups of countries
Factor Analysis (FA) weighting to create a composite index, based on four groups
(A, B, C and D) of basic indicators. The groups of basic indicators that were
considered refer to the three types of benchmarking we use: road safety
performance (final and intermediate outcome), policy performance indicators and
ckground indicators (motorization level, population density). The analysis used
the data collected for 27 European countries.

e trials of creating a composite index were performed, where each trial pro F
a
40
with similar values of the combined index. The composite index enables us to rank
the countries in accordance with their safety performance.

The composite indices and clusters/groups of the countries are further compared
among the trials.
3.5.1. Comparisons of countries' rankings
First, the weighted safety indices obtained with each analysis were considered
together (Figure 3.5), where the basic countries' ranking is built using their final
scores from the PCA-all analysis. It can be seen that:
The results of the 'PCA-groups' analysis are very different from all other results.
The results of the 'FA-2Factors' and the 'FA-2Factors-noC4' analyses are very
close.
In general, there is a certain similarity between the results of 'PCA-all', 'FA-
4Factors', 'FA-2Factors' and 'FA-2Factors-noC4' analyses.

d

dicators were initially analysed in the pre-defined groups, not accounting for the
ter-group correlations. Such a consideration enables a deeper insight into the
The different character of the results obtained with the 'PCA-groups' metho
probably stems from the different approach undertaken in this case, where the
in
in
behaviour of indicators in each group, in comparison to other analyses, but can
provide a different final picture when a combined index is created.

Final safety indicators from different analyses
1
2
3
4
W
F
PCA-all PCA-groups FA-4factors
FA-2factors-noC4 FA-2factors
-2
0
SE UK FR LU DE NO IE SI FI NL CH DK AT EE ES CY BE SK PT MT LV LT EL PL HU CZ IT
-1

Figure 3.5. Final safety indicators as a result of five trials.


Th ur of group
the
A,
pa
the next country rank 2 and so

e 'PCA-groups' analysis indicated a lack of similarity in the behavio
factors. Therefore, it was interesting to compare the results of other analyses with
intermediate results of the 'PCA-groups' analysis, i.e. with the factors built for the
B, C groups of indicators. Due to scaling reasons, prior to this comparison, we
ssed from the values of weighted indices to country ranks (where a country with
best combined safety indicator receives rank 1, the
on).
41
Fig
the nks based on the FA (A-group factor), FB (final B-group factor), FC
inal C-group factor) and FC1 (factor 1 from the C-group analysis, which reflects
ead of FC creates some differences in the
countries' ranking.
er of countries for which the ranks are consistent across the

to their safety level are Luxembourg, Slovenia, Malta and Italy.
ure 3.6 presents a comparison of country ranks from the PCA-all analysis with
country ra
(f
seatbelt wearing rates and median age of passenger cars, only) values. The PCA-all
ranks in Figure 3.6 are accompanied by 5 rank deviation bars (vertical grey lines).
It can be seen that:
The deviations from the PCA-all ranks are generally wider when the countries are
ranked in accordance with FA or FB factors than with factors FC or FC1.
Consideration of FC1 only inst
We can note a numb
different considerations. These are: Sweden, Great Britain, France, Germany,
Norway, Austria, Spain, Lithuania and Poland;
The countries with wide deviations in ranks and therefore the most unstable
answers as

Country ranks from different analyses
35
-10
-5
0
5
10
15
20
25
30
SE UK FR LU DE NO IE SI FI NL CH DK AT EE ES CY BE SK PT MT LV LT EL PL HU CZ IT
R
a
n
k
PCA-all ranks FA ranks FB ranks FC ranks FC1 ranks

Figure 3.6. Countries' ranks resulting from the PCA-all analysis and group factors
analyses.


Figure 3.7 presents another comparison of country ranks from the PCA-all analysis
nd separate factor analyses, where the basic countries' ranking is built using a
results of other rankings. In our case, as a basis for
itional approach, i.e. FB1 - factor 1
B-indicators together (fatalities per
a
traditional approach of countries' comparison in terms of fatality rates (per
population, vehicles, etc.). Such a comparison was performed, for example, by
Hermans et al. (2008), who used the countries' ranking based on personal safety as
a basis for judging the
comparisons, we applied a substitute to the trad
built in B-group analysis, which reflects five
population, vehicles, passenger km-travelled, injury accidents per fatality and share
of pedestrian fatalities). The countries' FB1 ranks are compared with PCA-all ranks,
FB (final B-group factor) ranks and FC (final C-group factor) ranks. The FB1 ranks in
Figure 3.7 are accompanied by 5 rank deviation bars (vertical grey lines).

42
Country ranks from different analyses
15
20
25
30
35
k
PCA-all ranks FB ranks FC ranks FB1 ranks
0
5
10
R
a
n
-5
-10
NL DE CH MT SE UK NO FI AT IT DK FR LU BE PT SI ES IE CZ CY SK EL HU EE PL LT LV

Figure 3.7. Countries' ranks resulting from the FB1 ranks versus PCA-all, FC and FB
ks. r

an

It can be seen from Figure 3.7 that:
For the majority of countries, the total FB ranks are close to the FB1 ranks, which
confirms the leading role of FB1 indicators (fatalities per head of population,
vehicles, passenger kilometres travelled, injury accidents per fatality, and share
of pedestrian fatalities) in countries' estimation by 'safety outcomes'.
Both PCA-all ranks and FC ranks create deviations from the countries' ranks
based on safety outcomes only. The countries with the widest deviations in ranks
and therefore with the most unstable answers in this context are Malta and Italy.
Essential deviations from the FB1 ranking are also observed for France,
Luxembourg, Slovenia, Ireland, Estonia and Latvia.
At the same time, we can note a number of countries for which the ranks are
relatively consistent across the different considerations. These are: Sweden,
Great Britain, Norway, Portugal, Spain, Slovakia, Greece, Hungary, Poland and
Lithuania.

Finally, to compare the countries' ranks based on the combined indicators, we
decided to present the results of four analyses, i.e. PCA-all, FA-4Factors, FA-
2Factors, and FA-2Factors-noC4 (where the results of the PCA-groups analysis are
left out due to their different nature), see Figure 3.8. As a basic ranking we chose
the one that provided the minimum sum of squared deviations from other rankings,
i.e. the FA-2Factors-noC4 ranking.

Figure 3.8 shows that:
Similar to countries' ranking based on the weighted safety indices (see Figure
3.5), the results of FA-4Factors, FA-2Factors and FA-2Factors-noC4 analyses
are reasonably close, whereas PCA-all ranks are associated with large
deviations.
Essential differences between all rankings are observed for Malta.
Relatively large deviations between some rankings are observed for the
Netherlands, Switzerland, Norway, Luxembourg, Ireland, Slovenia, the Czech
Republic, Italy and Estonia.

43
Final r a nt
0
5
10
15
20
25
SE UK DE R MT N I DK L IE ES BE CY PT EL CZ I K LV
R
a
n
k
anks ccording to differe analyses
30
NL CH F O F U AT SI T S EE HU PL LT
PCA-all-ranks FA-2f tors-noC4-ranks ac
FA-4facto nks rs-ra FA- tors-ranks

2fac
Figure 3.8. Final countries' ranks resulting from four analyse

s.

In general, the level of nsist cy between the different ra ng se s mediocre,
whereas the most similar results ar provided by a simila y o n sis (e
common factor analysis).
3.5.2. Identification of groups of countries
Due to various reason (e.g. data quality, method of analysis, random variations in
data), the country rankings from different analyses will prob ly ev b identic
Therefore, a more reasonable comp ison of th esults of v l a es cou
be made by forming groups of cou ies with similar safet pe r n . Such
comparison is possible clusters of
countries with r values of the combined indicator.

Using the res of c tries luste g presen in Sect and Appendix 2,
we chose to compare the countries' classifications into
summarizes the findings. The left part of Table 3.6 illustrate h o try roupin
obtained from e four nalys with ore cons ent resul i. FA F tors, F
2Factors, FA P A-all, wh as the h a o he tab
illustrates the country groupings obtained from the initial and final steps of the PCA-
ed on PCA-group', give the final classifications of the countries.
able 3.6 clearly shows that the countries' classifications in the left part are much
tab ), five groups of countries with different levels of safety
1. e highest level of safety performance: Sweden, Norway, France,
Great Britain, Germany;
2. Countries with a relatively high level of safety performance: Switzerland, the
Netherlands, Finland, Denmark, Ireland, Austria, Luxembourg, Malta;
3. Countries with a medium level of safety performance: Cyprus, Slovenia, Portugal,
Belgium, Spain;

co en nki s em
e r t pe f a aly .g.
s
ab n er e al.
ar e r se era an lys ld
ntr y rfo ma ce a
in our case, because each analysis produced
very simila
ults oun ' c rin ted ion 3.4
five groups. Table 3.6
s t e c un g gs
th a es m ist ts, e. -4 ac A-
-2Factors-noC4 and C ere rig t p rt f t le
groups' analysis. The two central columns, i.e. 'Final group: based on four analyses'
and 'Final group: bas

T
more consistent than those in the right part of the table. Based on the left part of the
le (results of four analyses
performance can be defined as follows:
Countries with th
44
Country FA- FA- PCA-
Final group:
based on four
Final group:
based on
PCA-group FA FB FC
WF
(combined
FA-
2factors-
analyses code 4factors noC4 2factors all index)
SE 1 1 1 1 1 1 1 1 1 1
NO 1 1 1 2 1 1 2 1 1 1
FR 1 1 1 1 1 1 1 1 1 1
UK 1 1 1 1 1 2 1 1 2 2
DE 1 1 1 2 1 2 2 1 2 2
CH 1 1 1 3 2 2 1 1 3 2
NL 1 1 1 3 2 2 1 1 2 2
FI 2 2 2 3 2 2 3 2 2 1
DK 2 2 2 3 2 2 3 2 2 1
IE 2 2 2 2 2 2 3 2 3 1
AT 2 2 2 3 2 3 3 1 4 2
LU 2 2 2 2 2 3 4 2 4 3
MT 2 1 1 4 2 3 3 1 3 4
CY 3 3 3 4 3 3 3 2 3 2
SI 3 3 3 3 3 2 4 2 2 2
PT 3 3 3 4 3 3 4 2 3 2
BE 3 3 3 4 3 3 3 2 5 3
ES 3 3 3 4 3 3 3 2 4 2
EE 4 4 4 3 4 3 3 4 3 1
SK 5 4 4 4 4 3 4 3 4 1
EL 4 4 4 5 4 3 3 2 4 2
CZ 4 4 4 5 4 4 4 3 5 2
LV 5 5 5 4 5 3 4 4 3 1
HU 5 4 4 5 5 3 3 4 5 1
PL 5 4 4 5 5 4 4 4 5 2
LT 5 5 5 5 5 4 4 4 5 2
IT 5 4 4 5 5 4 5 1 5 3
Table 3.6. Groups of countries with similar safety performance, based on the results
of various analyses.


4. Countries with a relatively low level of safety performance: Estonia, Slovakia,
Greece, Czech Republic;
5. Countries with a low level of safety performance: Latvia, Hungary, Poland,
Lithuania, Italy.
45
3.5.3. Behaviour of basic indicators
In the analyses performed, similarities in behaviour were observed for the following
passenger cars),
of
otorcyclist fatalities) and C7 (percentage of HGV in fleet) were excluded from
several analyses due to low correlation with other variables.

Based on the analyses performed, basic indicators with more consistent behaviour
and a clearer contribution to the final composite index can be identified. Such
indicators are: B1-B2-B3 (fatality rates), B5 (share of pedestrian fatalities), A1-A2-
A3-A4-A5 (quality of national safety programs), C2-C3 (wearing rates of safety belts)
and C5 (median age of cars). These indicators can serve as a core set of basic
indicators for the characteristic of a country's safety performance.
3.5.4. Towards the SUNflower road safety performance index
The purpose of our analysis was to create a composite road safety performance
index and, concurrently, to explore the similarities in basic indicators. It has been
demonstrated that both tasks can be realized by the statistical weighting methods
applied. The composite indices, estimated by several methods, enabled us to rank
the countries according to their safety performance.

Because of differences in rankings obtained by the different methods used, a more
reasonable comparison can be made by using groups of countries with similar levels
of safety performance. Countries can be grouped based on the values of composite
indices received. In particular, among the 27 European countries considered, five
groups with different safety performance were recognized. With the methodology
used and with the available data, the countries with the highest level of safety
performance are: Sweden, Norway, France, Great Britain and Germany. This group
remained fairly consistent among the different methods used. Although, of course, a
particular country could move from one group to another (neighbouring) group,
depending on the method used. The analysis revealed that the results of countries'
rankings based on the combined indicators are not necessarily identical to the
traditional ranking they receive based on mortality rates or fatality rates only.

We believe that adding the information on policy performance indicators and
implementation performance indicators to the ranking and grouping process
groups of indicators:

B1-B2-B3-B5-C5 (safety outcomes, i.e. fatality rates per population, per vehicles,
per passenger kilometres travelled; share of pedestrian fatalities, and the median
age of

A1-A2-A3-A4-A5 (indicators of the quality of national safety programs),

C2-C3 (wearing rates of safety belts in front and rear seats) and

D1-D2 (the country's background characteristics).

Moreover, the C4 indicator (average EuroNCAP score) was relatively close to the
group containing safety outcomes, where B4 (injury accidents per fatality) and C6
(percentage of motorcycles in fleet) were relatively close to the D1-D2 group. On the
other hand, C1 (share of drink-driving accidents) behaved different from all the other
indicators, whereas indicators B6 (share of bicyclists fatalities), B7 (share
m
46
improves the results from the establishe
comprehensible and meaningful.
d methods and makes them more

Moreover, it was observed that the indicators belonging to the final outcomes and
intermediate outcomes, both part of the road safety performance indicator, are not
uniform in their behaviour in the analyses. Indicators which were found to be more
consistent and named 'core set of basic indicators' are recommended for future
uses.

The overall conclusion is that it is realistic and meaningful to design a composite
road safety performance index in which relevant information from the different
components of the road safety pyramid has been captured and weighted. Moreover,
such an indicator gives a more enriched picture of road safety than a ranking only
based on data on mortality or fatality rates, which is normal practice today. Grouping
countries in this process is promising and seems to be preferable to simply ranking
countries. Country grouping will be done in the following chapter, Chapter 3. Before
efining the SUNflower road safety indicator and actually using the results for policy d
making, two improvements are recommended: firstly to design the Implementation
performance indicator and secondly to develop procedures for collecting high-quality
and comparable data for all three indicators for all EU Member States.

47
4. Grouping countries
It is important for countries to compare their safety performances with those of other
countries. A first motivation for comparison is to know how the overall safety
situation in the most recent year compares with that in other countries. Sometimes
the comparisons are expressed in terms of rankings. In order to do so, it is
necessary to define safety. Safety is often defined in terms of mortality rates:
with? This question is not easily answered. The
answer depends on the purpose of the comparison. If only a simple ranking of
time has to be carried out, then the situation is even more
d
large nu sy to make. Comparisons between a smaller range of
w
learn
conomica background, and/or the same level of
be considered as an extension of the Sunflower+6 study
hich used three groups of three countries each: the North-West European 'SUN'
countries (Sweden, the United Kingdom and the Netherlands), the Southern
fatalities per head of the population. Mortality rates are used primarily to rank traffic
safety or traffic risk to other risks, such as mortality due to diseases, during labour
accidents, or accidents in and around the house. For the comparison of traffic risks
this has the disadvantage that the degree of motorization is not taken into account.
Therefore, another indicator is commonly used as a criterion for traffic safety: fatality
risk, defined as the number of fatalities per motor vehicle kilometre. For those
countries in which the motor vehicle kilometres are not available, the fatality rate -
defined as the number of fatalities per motor vehicle - will be used instead.

Not only the recent safety situation is of interest, but also the safety development
over time: has the countrys safety been increasing or decreasing over time?
Therefore, trend analyses will be carried out in this chapter to enable comparisons
between countries over time.

A third reason for comparison is to learn from other countries. How can we improve
the safety situation in our country? Are there useful examples of safety policies in
other countries that can be identified? In this case more detailed information is
necessary.

For all three types of comparison the most important question is: which country do
we want to compare ourselves
countries in a certain year is required, then the fatality risk indicator seems to be
clear enough. However, even then it is not fair to carry out a direct comparison
between all countries. Some countries have a more difficult task to fulfil than others
and a correction for such a handicap should be applied. However, it is not easy to
measure and quantify such a handicap.

If a comparison over
complicate . In that case there is not one single indicator that unambiguously ranks
countries. It is not easy to define the-best-in-class this way. Nor is a comparison of
mber of trends ea a
countries ith similar traffic systems or safety levels, or with a more general
common background seem to be more promising. Of course countries can
from measures taken in all other countries. But to formulate targets or plans it is
more realistic to compare with countries in the same situation, and/or with the same
l, historical and geographical e
motorization and safety development.

Therefore it was decided to research the possibilities and consequences of grouping
European countries in such a manner that comparisons can be made against more
similar backgrounds. This chapter identifies three alternative ways to carry out such
a grouping. The first can
w
48
European countries (Spain, Portugal and Greece) and the Central European
countries (the Czech Republic, Hungary and Slovenia). The second grouping is
based on the level of safety and the safety development in the European countries
for which data is available. The third grouping is done on a much larger number of
social, economic and geographic characteristics. The first grouping differs from the
other two, because it is not directly based on objective data. Other than the number
f characteristics, the last grouping differs from the second one, in that it is based on
e analysis of fatality risks is highly preferable. There are ways of
stimating motor vehicle kilometres from the number of motor vehicles and fuel
e been tried here.
the first approach 13 variables which were readily available have been selected,
sed on the Sunflower+6 study
The first grouping was carried out by a small number of traffic safety experts based
of the countries
and their geographical position. The following grouping was suggested for the
European countries:
Group 1: Denmark, Finland, Iceland, Ireland, the Netherlands, Norway, Sweden,
o
the recent situation only and not on past developments.

Pilot studies of all three ways of grouping have been carried out and are compared
with each other in this chapter. These pilots are preliminary and carried out on
existing data that was easily available. The objective was primarily to show how the
existing techniques could be applied and how the results relate to each other. This is
also the case for the grouping that was finally chosen. For example, the first
grouping can be replaced by a study using a large number of experts, e.g. a study in
which a number of safety experts from different countries are asked to rank the
countries on a number of pre-selected characteristics. Techniques are available that
translate such similarity data into ranking of countries. The second way of grouping
is carried out using fatality risks (fatalities/motor vehicle kilometres) and fatality rates
(fatalities/motor vehicles) because vehicle kilometre data is lacking for many
countries. Th
e
sales. For the third method it turned out that a number of characteristics of countries
could not easily be collected, although this data will be available in some format in
various databases. Therefore, with this method a more decisive grouping is possible
than the groupings carried out in this study. Two approaches hav
In
including 3 road safety related variables. In the second approach the 3 road safety
related variables were excluded.
4.1. Grouping by safety experts ba
on the grouping in the Sunflower+6 project, the safety development

United Kingdom;

Group 2: Austria, Belgium, France, Germany, Italy, Luxemburg, Switzerland;

Group 3: Bulgaria, Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland,
Romania, Slovakia, Slovenia;

Group 4: Cyprus, Greece, Malta, Portugal, Spain.

There was some discussion about whether Italy should be placed in group 2 or
group 4, Ireland in group 1 or group 2, the Netherlands in group 1 or group 2 and
Slovenia in group 2 or 3.

49
In Section 4.2 this grouping based the opinion of road safety experts will be
compared with the grouping obtained with a Singular Value Decomposition (SVD) of
e annual fatality risks in the years 1980-2003. In Section 4.3 the grouping will be
haracteristics involved in the analysis.
mount of subjectivity involved in the grouping of countries based on
ese plots.
e technique.
4.2. composition of traffic safety
al, underlying, or
trend or dimension would be
th
done according to a Multiple Correspondence Analysis (MCA) of a number of
characteristics of the countries as observed in the most recent years; here their
fatality risks are just one of the c

The purpose of both the SVD and the MCA technique is to detect trends in
characteristics that countries have in common: that is, to capture the most important
and salient relations between the variables. In the SVD these are the developments
in fatality risk, in the MCA they are the characteristics of the countries (the
observational units) in just a few (e.g. two) dimensions. The main purpose of this
study is to capture the most important similarities between the countries. Both
techniques produce a plot in which each country is represented by a dot; in this plot
small distances between dots imply that the corresponding countries are quite
similar to each other (in terms of the values of the variables used in the analysis)
while large distances between dots represent countries that are very different from
each other. Analysis of the distances between countries in this plot is therefore
particularly suited for the grouping of countries, even though as we will see there
is still a certain a
th

As was mentioned earlier, a second result of these techniques is that they allow for
an assessment of the most salient relations between the variables used in the
analysis. In the context of the SVD of fatality risks of the countries, some attention
will therefore also be paid to this aspect of the analyses' results.

Whether to use an SVD or an MCA for our purposes typically depends on the
measurement level of the variables used in the analysis. When continuous
numerical variables (such as fatality risk) are used SVD is the most appropriate
technique. On the other hand, when we use discrete nominal variables, such as
religion, MCA is the appropriat
Singular Value De
developments
4.2.1. General description
One way of grouping countries is to look for similarities in the risk level and the risk
development. To investigate these similarities between European countries, a
Singular Value Decomposition (SVD) of the developments in fatality risks was
performed. A detailed description of this technique is given in Appendix 5. The SVD
looks for common trends in the risk development. The original trends for each
individual country will then be replaced by a small number of gener
latent trends, in such a manner that the individual trends are best represented by
this small number of general trends. The result is that for each country a
combination of general trends is found with a minimum of deviations from this
combination. If, except for a multiplicative factor, all countries had the same
development of risk over time, then one common
found, with a factor representing the level of risk for each country. This is often
called a factor score: the score of a country on a dimension. A low average risk will
then be represented by a low risk factor, a high level by a high factor. If we plot
50
these values on a straight line, then countries with similar levels of risk will be close
together.

In general, countries will differ not only in the level of risk, but also in the risk
development. That means that there are more general trends, and that the SVD will
find more than one dimension.

The first dimension will be the best representation of the development for all the
ountries. It is a weighted average of all risk developments. The next dimension is
the best representation of the r m im is
represents the most important correction of the first general trend. For example, if a
large number of countries have exa same correction over time, then their
deviation from the general trend will be made visible by this second dimension. For
each country a factor is found that represents the degree to which this trend is
important for that country. If the factor is almost zero, then the trend is not relevant
for that country. If the factor is positive, then the trend is relevant for that country, if it
is negative then the reverse of the trend is relevant for that country. If the factors for
the first and second dimension are plotted in a plane, the countries that have the
same general level of risk and the same deviation from the general trend will be
close together.

In principle, if the number of countries is lower than the number of years, there are
a factors as there are countr the s less important factors
represent less variation in the risk developments. Each dimension has a so-calle
eigenvalue, representing the importance of that dimension for the description of the
risk developments. The eigenvalues are a decrea ries. eneral trends
are found, only white noise remains, and the dimensions representing the white
noise will all have comparably small dimensions can then be
ignored. The choice of a cut-off value for the number of dimensions is somewhat
arbitrary. However, this study will focus on the first two or three dimensions to


European countries
ands have the lowest average risk over the years, followed by
c
esiduals fro the first d ension. Th dimension
ctly the
s many ies. At ame time,
d
sing se If all g
eigenvalues. These
represent the similarities between countries in risk development, even if more
dimensions show real trends.
4.2.2. Fatality risk developments (fatalities per motor vehicle kilometre) in 13
A first analysis was performed on the fatality risk developments. The fatality risk is
defined as the total number of fatalities in a country in one year divided by the total
number of motor vehicle kilometres for that year. The IRTAD data was used,
supplemented with additional information available from the Sunflower+6 project.
Data of motor vehicle kilometres for a long enough series of years was available for
only 13 countries. Data of the years 1980 through 2003 was used for the following
countries: Austria, Belgium, Denmark, Finland, France, Germany, Great Britain, the
Netherlands, Norway, Portugal, Slovenia, Sweden and Switzerland. For some
countries one or two values were missing and hence interpolated.

Five dimensions were used in the analysis. Four of these represented some trend or
trend correction; the last one already represented white noise. Table 4.1 shows the
country factors, the eigenvalues and the cumulative proportion of explained variance
in the total matrix of scores for the four dimensions.

Table 4.1 shows that in the first dimension Sweden, Great Britain, Norway, Finland
and the Netherl
51
Denmark, Switzerland and Germany. Portugal and Slovenia have the highest
s that
which
data matrix.
his shows clearly that all countries generally have the same risk development.
igure 4.1 shows this general trend in combination with the best exponential fit. It
turns out that there are two major deviations from the general smooth trend: one
around 1983 and one around 1987.

average risk, followed by Austria, Belgium and France. The table also illustrate
the first dimension is by far the most important one. The eigenvalue is 0.5955,
eans that the first dimension explains 99% of the variance in the total m
T
F
Country Dimension 1 Dimension 2 Dimension 3 Dimension 4
Austria 0.2598 -0.0949 0.0618 0.7291
Belgium 0.2335 0.0794 0.3644 0.0173
Denmark 0.1453 0.1916 0.2002 -0.2899
Finland 0.1216 0.2107 0.1635 -0.2911
France 0.2213 0.1588 0.3263 0.0643
Germany 0.1741 0.0943 0.2037 0.3614
Great Britain 0.1090 0.0404 0.1684 0.0509
The Netherlands 0.1262 0.0718 0.1990 0.0600
Norway 0.1116 0.2047 0.2553 -0.0877
Portugal 0.6205 -0.7205 -0.0276 -0.2817
Slovenia 0.5529 0.4977 -0.6557 -0.0050
Sweden 0.0951 0.2016 0.2175 -0.2640
Switzerland 0.1512 0.1153 0.1937 0.0410
Eigenvalue 0.5955 0.0343 0.0297 0.0188
Cumulative proportion of variation 0.99279 0.99608 0.99856 0.99955
Table 4.1. Country factors, eigenvalues and the cumulative proportion of explained
variance in the total matrix of scores for four dimensions from an SVD analysis of
fatality risks (fatalities per 10
6
motor vehicle kilometres) for 13 European countries.


The second and third dimensions have much smaller eigenvalues than the first
dimension, but they do not differ much. Figure 4.2 shows the second and third
d by
e negative value for Portugal means that the correction on
e general trend is reversed. This means that Portugal is mainly responsible for the

r on these dimensions, and especially for those countries with a
dimension and a weighted combination of these two dimensions. The bold figures in
the third column of Table 4.1 show that the second dimension is dominate
Portugal and Slovenia. Th
th
peak in 1987, and that especially Slovenia has a low value in that year. The third
dimension is negatively dominated by Slovenia and positively by Belgium, France
and Norway. This means that Slovenia is primarily responsible for the peak in 1983.
Norway also had a rather high positive score on the second dimension, which
indicates that especially Norway has a relatively low value in 1983. Both correction
trends increase with time, and are more or less mirrored. The weighted average of
the trends shows a rather linear correction which increases with time. For countries
with a positive facto
positive factor on both, such as Norway, Denmark, Finland, Sweden and France,
this means that the development over time decreases a bit less than the general
trend suggests.

52
General trend in fatality risks in 13 countries
0,2
0,3
d

f
a
c
t
o
r
0
0,1
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002

0,4
T
r
e
n
Figure 4.1. General trend (first dimension) in fatality risks, resulting from SVD
analysis of 13 countries, together with the best fitting exponential trend.


Trend in second and third dimension
-0,6
4
-0,2
0
0,2
0,4
0,6
1980 1982 1984 1986 1988 1992 19 1998 02
T
r
e
n
d

f
a
c
t
o
-0,
1990 94 1996 2000 20
r
Dimension 2 Dimension 3 Average

Figure 4.2. Second and third dimension (residuals), and their weighted average,
resulting from the SVD analysis of the fatality risks untrie

of 13 co s.

Slovenia has opposite factors, which suggests a relatively high risk in the early
1980s and early 1990s, a relatively lo e late 1980s, an erage
the late 1990s and the early years of ond millennium. Th tive s
Portugal on the second dimension shows that there was a relatively high peak in risk
around 1987, a low risk during the first half of the 1990s and also, although to a
lesser amount, from the late 1990s onwards. The fourth dimension is dominated by
Austria (showing a relative high risk in the early 1980s, compared to a relatively low
risk in the late 1980s). Figure 4.3 shows the factor scores on dimensions 1 and 2
from Table 4.1 plotted f untries.

w risk in th d an av risk at
the sec e nega core for
or the 13 co
53
Factor scores of SVD analysis of fatality risks
0,6
-0,8
-0,6
-0,4
-0,2
0
0,2
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7
Dimension 1
D
i
m
e
n
s
i
o
n

2
0,4
SE
UK
NL
CH
DE
FR
BE
NO
FI
DK
AT
PT

SI
Figure 4.3. Factor scores for 13 countries on the first two dimensions of a SVD
analysis on fatality risks.


Based on Figure 4.3, the following grouping of countries can be made: in the upper-
left part of the plane the Nordic countries together with Great Britain and the
Netherlands are located. To the right of this group and a bit lower Switzerland,
Germany, France, Belgium and, a bit further away, Austria can be viewed as a
second group. Ellipses have been drawn only to indicate these two groups,
suggesting a possible grouping of these countries. Slovenia and Portugal have
isolated positions.

As expected, the countries with a low risk level - Sweden, Great Britain, Norway and
the Netherlands - are located at the left of the first dimension; Slovenia and Portugal
at the right.
4.2.3. Fatality rate developments (fatalities per number of motor vehicles) in
20 European countries
To investigate the similarities between more countries, an SVD analysis was
performed of the fatality rates (fatalities divided by the number of motor vehicles) for
s list:
, Hungary, Iceland, Ireland, Italy, Luxembourg and Spain.
ot enough data was available from among others Poland, Slovenia and Slovakia.
atality rates from 1980 through 2005 were used. The results are given in Table 4.2.

rst
tain, the Netherlands and Switzerland,
llowed by Italy, Germany and Switzerland. The countries with the highest factor
each country for each year. The following countries were added to the previou
Czech Republic, Greece
N
F
Although a bit less than in the analysis of fatality risks, the first dimension is once
more the dominant one. The countries with the lowest factor scores on the fi
dimension are: Norway, Sweden, Great Bri
fo
scores on the first dimension are Portugal, Greece and Hungary. The second
dimension has the highest positive factor scores for the Czech Republic and
Hungary and to a lesser extent for Portugal. Austria and Greece have the largest
negative factor scores. The third dimension is dominated by Norway, which has a
positive factor score. Second is the Czech Republic with a somewhat smaller
54
negative factor score. On the fourth dimension the Czech Republic has a high
positive score while Portugal and Hungary have large negative scores.

Country Dimension 1 Dimension 2 Dimension 3 Dimension 4
Austria 0.2239 -0.3332 0.0273 -0.0613
Belgium 0.2165 -0.0878 -0.1318 -0.0601
Czech Republic 0.1968 0.4825 -0.3756 0.6123
Denmark 0.1592 0.0493 -0.2234 -0.0744
Finland 0.1491 -0.0434 -0.1134 -0.1131
France 0.2168 -0.1314 -0.0915 0.0064
Germany 0.1442 -0.1774 0.0077 0.1385
Great Britain 0.1115 -0.0988 -0.0801 -0.0425
Greece 0.4027 -0.3045 -0.1493 0.1203
Hungary 0.3873 0.5736 -0.1046 -0.3905
Iceland 0.0996 -0.0096 -0.0863 0.0947
Ireland 0.2250 -0.0916 -0.0944 0.1644
Italy 0.1415 -0.1374 0.0248 0.1416
Luxemburg 0.1880 -0.1884 -0.0594 0.0540
The Netherlands 0.1325 -0.1437 -0.1601 -0.0182
Norway 0.0908 -0.0478 0.8005 0.2113
Portugal 0.4438 0.1828 0.0062 -0.5437
Spain 0.2430 0.0343 -0.1219 -0.0703
Sweden 0.0944 0.0095 -0.0070 -0.0198
Switzerland 0.1330 -0.2084 -0.1677 -0.0806
Eigenvalues 9.1632 0.7887 0.6246 0.4817
Cumulative proportion of variance 0.9826 0.9899 0.9944 0.9971
Table 4.2. Country factor scores, eigenvalues and the cumulative proportion of
imensions from a SVD
ehicles) for 20 European
that the peak in those years was underestimated for these countries. The
explained variance in the total matrix of scores for four d
analysis of fatality rates (fatalities per number of motor v
countries.


Figure 4.4 shows the first dimension of the SVD analysis on fatality rates. A sudden
rise in the fatality rate in the early 1990s can be observed. This rise is driven by the
Czech Republic, Hungary and Portugal, as can be seen in Figure 4.5, which
displays the graphs for the second, third and fourth dimension. The graph of
dimension 2 also shows a peak value in the early 1990s. From Table 4.2 it can be
seen that most countries have a negative factor, showing a correction on the rise in
dimension 1. Especially Hungary and the Czech Republic have high positive values,
howing s
third dimension has a similar peak as dimension 2, but shows a steeper decline
later. Except Norway, most countries again have a negative or neutral factor score.
This shows a more than average safety improvement in Norway from the early
1990s onward. Compared with Portugal and to a lesser extent with Hungary, the
positive factor for the Czech Republic on dimension 4 shows that the relatively
55
positive development until 1990 turned into a relatively negative trend in later years,
tending to neutral in most recent years.

General trend in fatality rates in 20 countries
0,4
0,2
T
r
e
n
d

f
a
c
t
o
r
0
0,1
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004
0,3

Figure 4.4. General trend (first dimension) in fatality rates, resulting from SVD
nalysis of 20 countries. a


Trend in second, third and fourth dimension
0,4
0,6
-0,2
0,2
0
d

f
a
c
t
o
r
T
r
e
n
-0,4
-0,6
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004
Dimension 2 Dimension 3 Dimension 4

Figure 4.5. Second, third and fourth dimension (residuals) resulting from the SVD
analysis of the fatality rates of 20 countries.


re 4.6 shows the factor scores on the first and second Figu dimension. Compared
with Figure 4.3 we see that Germany has slightly moved in the direction of the
Netherlands. It can also be observed that the development in Iceland is similar to
that in the Nordic countries and that Luxembourg is rather close to France and
Belgium. Ireland is close to Belgium and France; Spain is in between this last group
and Portugal, although closer to Belgium and France. Hungary and the Czech
56
Republic are close to each other on the second dimension, but not on the first.
Greece is a bit isolated, but closest to Portugal and Spain.

Factor scores of SVD analysis of fatality rates
0,4
0,6
HU
CZ
-0,4
-0,2
0
0,2
0
D
i
m
e
n
s
i
o
SE
NO
UK
DK
NL
LU
BE
FR
DE
IS
IE
CH
0,1 0,2 0,3 0,4 0,5
n

2
AT
ES
EL
IT
PT
FI
Dimension 1

Figure 4.6. Factor scores for 20 countries on the first two dimensions of a SVD
analysis of fatality rates.


Considering the grouping of countries, it can be concluded that the grouping as
result of the SVD analysis of fatality risk and rate trends is largely in agreement with
the grouping by the experts. The group 1 result for the Nordic countries, together
with Iceland, the Netherlands and Great Britain agrees with the expert selection for
group 1. Italy seems more similar to the countries in group 2 than to the Southern
countries Spain, Portugal and Greece. It also seems best to add Ireland to group 2.
The Czech Republic, Hungary and Slovenia can be grouped together in group 3.
4.3. Multiple Correspondence Analys
The idea behind this approach is that there are many factors other than traffic and
s hat might influence the system possi
f ples are the geographical, social or cultural charac-
t iety. To compare a y's safe ation with
o , it is convenient to select those tries that are similar regarding
these backg is a great variety of characteristics and they are
n easurable in the ordinary sense. Examples are type of religion,
l native to the Principal Co nts Ana (PCA) tech
w this type of data is the Multiple Correspondence Analysis
(MCA).

B of a different nature and are not directly related to traffic
safety, it is difficult to collect the necessary information for all ean countr
o how such a technique can be for the purpose of gro
countries, a small number of readily available characteristics were used in the
analysis discussed in this section. A disadvantage of the choice that was made is
t f the used characteristics the same type or are strongly
correlated with other characteristics. Examples are variables ter tempera
is
afety characteristics t traffic or restrict bilities
or safety measures. Exam
eristics of a country or soc countr ty situ that in
ther countries
round variables. There
coun
ot all directly m
iteracy etc. An alter mpone lysis nique
hich is suitable for
ecause the data are
Europ ies. In
rder to show used uping
he fact that many o are of
6 (win ture),
57
7 (summer temperature), 9 (latitude) and 10 (religion), which are defined below. The
outcome of any MCA analysis will therefore be dominated by such clusters, because
f
h artefacts disappear. Despite its drawbacks, the analysis is still
he data used in the analysis was collected for 23 countries: Austria, Belgium,
Czech Republic, Denmark, Finland, France, Germany, Great Britain, Greece,
,
he technique considers all variables to be measured at a nominal level; this means
rowth between 2000-2005): less than 1.75%;
.
2 ity rate per ve et (mean kille ousand motor vehicles between
005): less th 0; 0.120-0.16 -0.205; 0.206-0.305; higher than
.
3 te per po (mean killed ,000 inhabit een 2000-
s than 7. -12; 12.5-13.5; higher than 13
4 (aver three road types): less than 80%; 80-91%; higher than
5 in area (percentage of mountainous area): less than 5%; 5-30%; more
.

less than 18C; 18-19 C; 19.5-22C; more
than 22C.
8. Part of population with at least upper secondary education: less than 75%; more
than 75%.
9. Latitude (of capital, in degrees): less than 48; 48-50; 51-53; more than 53
10. Main religion: catholic; protestant; orthodox; secular.
11. Literacy (in population): less than 99%; 99%; 100%.
12. Population density (in absolute numbers per square km): less than 25; 25-99;
100-120; 122-180; more than 180.
13. GDP: gross domestic product (in million Euros): less than 90,000; 99,500-
210,000; 220,000-310,000; more than 400,000.

The classification results in a matrix of 23 countries by 13 variables. The scores
range from 1 to a maximum of 5: e.g., 1 or 2 for education, 1, 2 or 3 for literacy and
1-5 for growth of the vehicle fleet. The MCA of this data set was performed in two
dimensions. The dimensions are abstract characteristics, representing combinations
of the original variables that best represent similarities between countries. The first
dimension is again the most important common trait of the countries. The second

these are shown as dominant underlying components. In a more balanced set o
variables suc
reported here, in order to show the usefulness of the technique for grouping
countries.

T
Hungary, Iceland, Ireland, Italy, Luxemburg, the Netherlands, Norway, Poland
Portugal, Slovakia, Slovenia, Spain, Sweden, and Switzerland.

T
that the categories of a variable are just labels, without any order in the categories.
Thirteen variables were used. The values of the countries on these variables, if
numeric, were divided into a small number of classes. The classes were defined in
such a way that each category has an approximately equal number of countries.

The following discretized categories of the original variables were used in the MCA
2
:

. Growth of vehicle fleet (mean g 1
1.75-2.15%; 2.3-2.99%; 3-4.2%; more than 4.5%
. Fatal hicle fle d per th
2000-2 an 0.12 0; 0.165
0.305
. Fatality ra
): les
pulation per 100 ants be
.5.
tw
2005 5; 7.5-10; 11
. Seatbelt use age of
91%.
. Mounta
than 30%
6. Winter (mean January) temperature: less than 0C; 0-1 C; 1.5-4C; more than
4C.
. Summer (mean July) temperature : 7

2
The categories are in increasing order (1, 2, )
58
dimension is the next important one, describing best the residuals that cannot be
described by the first dimension. The two dimensions are independent from each
other. To what extent all information in the original matrix of 23 x 13 scores can be
isplayed in the two dimensions can be measured by the total fit. It turns out that the
total fit is equal to 0.853, which should be compared to a maximum fit of 2 (the
number of dimensions). In order to understand the meaning of the dimensions, the
so-called discrimination measures give an indication.

The major outcomes in terms of the variables are given in Table 4.3. In this table the
discrimination measures of the variables are given. The variables with the largest
discrimination measures on the first dimension are best represented by this
dimension. They are: main religion, summer temperature, population density,
latitude, fatality rate per vehicle fleet (fatalities/1000 vehicles), winter temperature
and fatality rate per population (fatalities/100,000 inhabitants). Growth of vehicle
fleet, seatbelt use, mountain areas, education and GDP are least represented by the
first dimension.

For the second dimension the most discriminating variables are: fatality rate per
vehicle fleet, summer temperature, latitude and population density. The least
discriminating variables are: growth of vehicle fleet, fatality rate per population,
seatbelt use, mountain areas, winter temperature, education main religion, literacy
and GDP.
are given in the last column:
ummer temperature, fatality rate per vehicle fleet, latitude and population density;
that order. Least discriminating are: mountain area, growth of vehicle fleet, GDP,
d

The most discriminating variables on both dimensions
s
in
education and literacy. According to the MCA this means that, given the selection of
variables, the 23 countries can best be grouped on the basis of the most
discriminating variables. These variables discriminate best between (groups of)
countries.

Dimension
Variable
1 2
Mean
Growth of vehicle fleet (5) 0.208 0.134 0.171
Fatality rate per vehicle fleet (5) 0.633 0.882 0.758
Fatality rate per population (5) 0.617 0.251 0.434
Seatbelt use (3) 0.209 0.266 0.238
Mountain areas (3) 0.214 0.055 0.135
Winter temperature (4) 0.620 0.291 0.456
Summer temperature (4) 0.741 0.814 0.778
Part of population with at least upper
secondary education (2)
0.254 0.293 0.274
Latitude (4) 0.687 0.647 0.667
Main religion (4) 0.743 0.254 0.498
Literacy (3) 0.489 0.079 0.284
Population density (5) 0.694 0.620 0.657
Gross domestic product (4) 0.102 0.297 0.199
Total 6.210 4.885 5.548
59
Table 4.3. Discriminati ures of the 1 les from a MCA analysis in two
dim .
Figure 4.7 displays the outcomes for the countries on the two d ns. Also in
t untries th close together have more in common th s
that are far apart.

on meas 3 variab
ensions
imensio
an countrie his case, co at are
Country pl
3
ot using 13 characteristics
-2
-1
0
1
2
-3 -2 -1 0 1 2
Dimension 1
D
i
m
e
n
s
i
o
n

2
SK
PL
CZ
LU
HU
AT
FR DK
NO
SE
FI
CH
DE
IS
NL
UK
IE
BE
SI
IT
PT
ES
EL

Figure 4.7. Plot of the 23 countries in the two-dimensional plane, resulting from the
MCA, using 13 characteristics of these countries.


Four groups of countries can be distinguished. Table 4.4 shows these groups. The
countries marked with * are similarly grouped as with the earlier two methods. Group
1 and group 4 agree to a large extent with the previous grouping by the experts and
the SVD analysis. Group 2 seems to be a mix of the original group 2 and group 3.
The new group 3 represents the countries in the centre of the plot, countries that are
least represented by the two dimensions. A third dimension could possibly show
how to best classify these countries.

Group 1 Group 2 Group 3 Group 4
Denmark*
Finland*
Germany
Great Britain*
Iceland*
The Netherlands*
Norway*
Sweden*
Switzerland
Austria*
Czech Republic
France*
Hungary
Luxembourg*
Poland
Slovakia

Belgium
Ireland
Slovenia
Greece*
Italy
Portugal*
Spain*
Table 4.4. Grouping of 23 countries in four groups resulting from an MCA using 13
characteristics.


60
We will end this section by presenting th
but now after removing the three main roa
e results of an MCA of the same data set,
d safety variables growth of vehicle fleet,
fatality rate vehicles', and fatality rate population from the analysis, see Figure 4.8.

Country plot using 10 characteristics
2
-1
0
D
i
m
e
n
s
i
o
n

2
ES
PT
IT
IE
SI
1
AT
CH
HU
PL
CZ
SE
LU
IS
NO
DK
NL
UK
FR
SK DE
BE
FI
-3
-2
-2 -1 0 1 2 3
Dimension 1
EL

Figure 4.8. Plot of the 23 countries in the two-dimensional plane, resulting from the
MCA, using only 10 characteristics of these countries.


There are some similarities between how the countries are clustered in Figures 4.7
and 4.8. Greece, Italy, Spain and Portugal, for example, still form a group in Figure
4.8. But there are also some large differences. For instance, other than in Figure
4.7, the Netherlands and Great Britain are not in the same cluster as Denmark,
Finland and Norway anymore, because the clustering is no longer based on the
similarities between these countries in terms of their respective road safety levels.
The most important variables responsible for the clustering found in Figure 4.8 are
(non-linear transformations of) the categories of latitude, population density, and
winter temperature. This illustrates the importance of a careful selection of the
variables used as the basis on which to form groups of countries.
4.4. Grouping countries, based on three strategies
Combining the results of the three grouping strategies discussed above, the
countries of which some time series data is available were classified into four groups
for further analysis. Countries for which annual fatality data and the number of
vehicles from 1970 onward were available were used in the analyses. Data from
980 onward was available for the group 3 countries Czech Republic and Hungary;
ctions 5.4 and 5.5.

1
this data was used for group 3. However, for Iceland, Poland, Slovakia and Slovenia
there was not enough data available for the annual number of vehicles in early years
to be used in most of the analyses. The final grouping of countries is given in Table
4.5. These groups are used for further analyses in Se
61
Group 1 Group 2 Group 3 Group 4
Denmark
Finland
Great Britain
Iceland
The Netherlands
Norway
Sweden
Austria
Belgium
Czech Republic
Hungary
Greece
Portugal
France
Germany
Ireland
Italy
Poland
Slovakia
Slovenia
Spain
Luxemburg
Switzerland

Table 4.5. Final grouping of countries on the basis of three grouping strategies.



62
5. Time series analysis
This chapter will first show a graph of the safety developments in 20 European
countries from 1970 onward (Section 5.1). In the following sections two different
approaches will be presented that can be used for time series analysis of safety
data: the SVD technique which was introduced in Chapter 4 and a special
application of the structural time series model developed by Harvey (1989).

In Chapter 4, SVD analyses were carried out to investigate whether similar
developments in fatality risk and rate could be used as a basis for grouping. Section
5.2 will continue with a more detailed study of the fatality risk development found
from the SVD analysis. It will mainly be concerned with the developments
themselves and not with the grouping of countries.

In Section 5.3 safety developments of individual countries are analysed to
investigate whether statistically unexpected safety outcomes are found in a certain
r with
ted
tality numbers, significant changes can be noted. However, it is better to use the
hole range of information about the fatalities as well as travel over a series of
arveys time series analysis technique. In this
tudy this has been done by modelling the development of fatalities as well as the
number of vehicles or vehicle kilometres. Then error bounds are not just based on
e
t.
. In
arveys structural time series model, observed fatality risk or rate outcomes are
nalysed with the SVD technique.
eneral trends for groups of countries are shown, together with deviations from
these trends for individual countries. Differences in outcomes for the different groups
are discussed.

Section 5.5 discusses the developments of disaggregate safety data for countries
within groups. Differences in age, traffic mode and road type will be presented for
each group. Some examples will be discussed in detail. All disaggregate
developments will be given in a separate appendix.
year. A crude way of doing so is to compare a safety outcome in a certain yea
the outcome of the previous year. Under the assumption of Poisson distribu
fa
w
previous years and not just information from the one previous year. An elegant way
of doing this is by making use of H
s
the number of fatalities in the previous year, but on the expected figures for th
amount of traffic and fatalities in the present year, given the outcomes in the pas
hen available, fatality risks are used, for the other countries the fatality rates W
H
compared with expected outcomes given the developments so far, together with
forecasts for the next three years. Error bounds are estimated for the safety data
over the whole range and also for the forecasts. This way it is possible to evaluate
annual fluctuations in safety risk or rate as well as trends, whether these are within
or beyond the margins of expectation. The expectation is determined based on the
number of vehicle kilometres or the traffic volume expected in a particular year and
on the expected risk or rate for that year, given the values in previous years.
Significant deviations in the number of fatalities can be caused by unexpected fleet
or traffic volume data as well as by unexpected changes in fatality risk or rate.
Harveys time series analysis technique distinguishes between these effects.

n Section 5.4 trends for groups of countries are a I
G
63
5.1. Fatality and mortality developments in 20 European
countries
Figure 5.1 shows the developments of the fatality and mortality rates in 20 European
countries from 1970 onward over five year periods. For some countries data is not
available over the entire period.

Figure 5.1 shows that in general all trends are the same, moving from the upper
right to the lower left. For the countries with a later start of mass motorization such
as the Czech Republic, Greece, Hungary, Poland, Portugal and Spain the beginning
of the curve is a bit chaotic.

Fatality rate vs. mortality rate for 20 European countries
0,3
0,45
0
0
0

i
n
h
a
b
i
t
a
n
t
s
Austria
Belgium
Czech Rep.
Denmark
0,35
0,4
)
Finland
France
Germany
Great Britain
0,1
0,15
0,2
0,25
M
o
r
t
a
l
i
t
y

r
a
t
e

(
f
a
t
a
l
i
t
i
e
s
/
1
Greece
Hungary
Ireland
Italy
Luxemburg
Netherlands
Norway
Portugal
Spain
Sweden
0,05
0
0 0,5 1 1,5 2 2,5
Fatality rate (fatalities/1000 vehicles)
Switzerland
Poland
Figure 5.1. Fatality rate vs. mortality rate in five year periods since 1970 for 20
European countries.


Generally, at the beginning of motorization, the mortality rate in particular can be
expected to rise with an increase of the number of vehicles, and to drop again later,
after an increased implementation of safety measures to counter the negative safety
effects. This is the case in Portugal, Greece and Hungary. For the Czech Republic,
the situation is different. Although the mortality rate rose steeply during the first
periods, there has been an only minor improvement in the fatality rate afterwards.
This causes a minor dip in the mortality rate. Compared with the other Central and
Eastern European countries the fatality rate was already rather low in the 1980s. For
the other countries this turning point lies before 1970, when the measurements
commenced. The general impression is that initially developments show a great
64
variety in values, to end within a narrow band. This suggests that, although
f safety generally remain
ahead of the other countries, albeit with a decreasing advantage.
mpared with the exponential trend, while the values for 1987 and
developments have different starting points in different countries, all countries end
up the same. However, countries leading in the field o
5.2. SVD analysis of the fatality risk developments in 11
European countries
In order to study the general fatality risk trend that was found in the SVD analysis of
13 countries with vehicle kilometre data in Section 4.2.2 more closely, an SVD
analysis was carried out, leaving out the two most disturbing countries, Portugal and
Slovenia. The result for the first dimension together with an exponential trend is
given in Figure 5.2. The percentage of variance explained by the first dimension is
now 99.37%. It is also clear that the exponential fit is much better. 99.21% of the
variation in the values on the first dimension is explained by the exponential function
Y = Exp(112.9638-0.0576*year) for the normalized fatality risks. The value for 1983
is relatively high co
1996 are relatively low. These deviating values are not explained by deviating motor
vehicle kilometres in those years.

General trend in fatality risks in 11 countries
0,4
0
0,2
0,3
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
T
r
e
n
d

f
a
c
t
o
r

0,1
Figure 5.2. General trend (first dimension) in fatality risks, resulting from SVD
analysis of 13 countries, together with the best fitting exponential trend.


Figure 5.3 shows the results for the 11 countries, if the first dimension is multiplied
by the factor score for each country.

65
Fatality risk estimates for 11 countries
0
0,01
0,02
0,03
0,04
0,05
1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002
E
s
t
i
m
a
t
e
d


f
a
t
a
l
i
t
y

r
i
s
k
0,06
Austria Belgium Denmark Finland France Germany
Great Britain Netherlands Norway Sweden Switzerland

Figure 5.3. Estimated fatality risks (fatalities per 1000 vehicle kilometres) from the
first dimension of an SVD analysis of the fatality risks in 11 European countries.


In Figure 5.4 the factor scores for the first SVD dimension of the 11 countries are
plotted, together with the fatality risk in the last year of measurement, 2003. This
figure shows that the overall decrease in risk from 1980 onward, measured by the
factor score, is linearly related to the risk in the last year of measurement, 2003.
Once more, this shows that the fatality risk development from 1980 onward largely
follows the same curve for all 11 countries, with a country-specific multiplicative
factor. This means that the absolute differences in fatality risk in the eleven
countries become smaller with time. The relative difference in fatality risk, however,
stays the same. If the fatality risk curves had been plotted on a log-scale, the curves
would have been parallel lines. This means that the leading countries in road safety
will keep their leading position if nothing changes.

Comparison of overall fatality risks of 11 countries
0,004
0,005
0,006
0,007
0,008
0,1
F
a
t
a
Sweden
Great Britain
Finland
Netherlands
0,009
0,012
0,013
0,014
0,15 0,2 0,25 0,3 0,35 0,4 0,45 0,5
Factor score
l
i
t
y

r
i
s
Norway
Switzerland
Belgium
0,010
0,011
k

2
0
0
3
Denmark
Germany
France
Austria

Figure 5.4. Comparison of overall fatality risk, expressed by factor score and the
fatality risk in 2003 (fatalities per 1000 vehicle kilometres) for 11 countries, together
with a linear trend.
66
5.3. Trends in fatality risks and rates for individual countries
s not available the annual figures for the total number of motor vehicles in a
ountry was used for this purpose instead.
5.3.1. Model stru
For each country a bivariate local linear trend model, called the latent risk model,
w d to estimate oscopic s and f sts for th elopments of
road safety and exposure. Whereas in a ssical regression model the intercept
and the regression co nt of th regression of a de nt variable in
time are fixed, and do not change o e, in a linear trend model these two
In this
l and
ires the estimation of 13 parameters: 3 for the
nce, yet another 3 for the disturbance variances of the
lope components of exposure and risk including their covariance, and, finally, 4
parameters for the initial values of the two level and the two slope components.

The model input data consists of the number of fatalities and vehicle kilometres per
year. The specific model structure is given in Appendix 6.
5.3.2. Results for three countries
In this section the outcomes of the time series analysis will be described in detail for
three countries. The results for all countries are given in Appendix 7.

Figure 5.5 shows the observed fatality numbers (marked with +) for France from
1970 till 2006, together with model values and error bounds. At the end of the
observation period forecasts are made for the next three years. It is seen that the
error bounds in the beginning are wide, due to a low value for the observed number
of fatalities in 1970 and the uncertainty of the model for the beginning of the period.
The error bounds become narrower at the end of the curve, probably due to the
smoother risk-curve. The error bounds are also wide for the forecast values,
over
cate the extent of random
uctuations. Such deviations should not be interpreted as significant changes.
For all European countries with available vehicle kilometre or vehicle data, the
Harvey structural time series model is used to analyse time trends for the number of
fatalities. The total annual number of road fatalities in each country was used as an
indicator for road safety. Whenever available the annual total number of motor
vehicle kilometres driven was used as an indicator for the exposure in a country; if
this wa
c
cture
as use macr trend oreca e dev
cla
ar efficie e line pende
ver tim local
parameters are typically allowed to change from time point to time point.
context the time-varying intercept and the regression weight are called the leve
he slope component, respectively. t

The latent risk model is a special kind of state space method for the analysis of time
series (Harvey, 1989; Durbin & Koopman, 2001; Commandeur & Koopman, 2007).
In the latent risk model the development of road safety is assumed to be the product
of the developments of two latent, unobserved factors: exposure and risk, see also
Bijleveld et al. (2008). The model requ
disturbance variances of the mobility and fatality figures including their covariance,
another 3 for the disturbance variances of the level components of exposure and
isk including their covaria r
s
probably because of the substantial jumps in the number of vehicle kilometres
the whole period.

Overall, the majority of observations are located within the error bounds. Differences
between observed values and the prediction line indi
fl
67
Initially, is no infor from
previous years. After a few years the error bounds becom r. Sometime
t ds bec ome se une d (s
changes in risk and/or m were observed.

the error bounds are rather wide because there mation
e narrowe
xpecte
s
udden he boun ome wider at s points, becau events
otorization)
1970 1975 1980 1985 1990 1995 2000 2005 2010
5000
7500
10000
12500
15000
17500
20000
Fatalities: France

Figure 5.5. Observed and predicted annual fatality numbers for France, using mo
vehicle kilometre figures a
tor
nd fatalities in Harvey's structural time series model.


1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1500
2000
2500
3000
3500
4000
Fatalities: Netherlands

Figure 5.6. Observed and predicted fatality numbers for the Netherlands, using
motor vehicle kilometre figures and fatalities in Harvey's structural time series
model.

nds
to the fact that there are fewer observations in a smaller
ountry. The curve is even more unstable at the beginning of the observation period.
ere the error bounds also become narrower at the end of the series. There are

Figure 5.6 shows a somewhat different result for the Netherlands. The error bou
re wider, only partly due a
c
H
68
three observations at the end of the observation period that are outside or just inside
the error bounds. The estimation curve shows a dip just before that period.
However, the three values are still extremely low. Therefore, this looks like a
ignificant drop in the number of fatalities, asking for an explanation. However, the
error bounds for the next three forecasted years remain reasonably narrow, despite
nd
igure 5.7 shows the observed and predicted fatality numbers for Greece. Because
fatalities. For some countries the increase in the number of fatalities is
efore 1975 and therefore not visible in the analysis from 1975 onwards. The model
can cope with an initial increase in fatalities, later followed by a consistent decrease,
because the basic developments used in the model are the vehicle kilometres or
fleet data and the fatality risks or rates, and these developments in principle show a
monotone increase and decrease respectively. The initially strong increase in the
number of vehicles and thus in vehicle kilometres levels off in later years, while the
fatality risk and rate decrease steadily. At the point where the exposure starts to
level off, the maximum number of fatalities is found. If these developments are
rather smooth, then the error bounds on the fatalities will not change. However, the
uncertainty that results from incidental jumps is reflected in the wider error bounds at
some of the periods and the very wide error bounds for the forecast period of three
years.

s
this drop. This indicates that the drop is interpreted by the program as part of a tre
and not as an incidental fluctuation.

F
motor vehicle kilometre data is missing, vehicle fleet data and fatalities are used.
This figure shows that the linear trend model is flexible enough to cope with
substantial changes in the shape of the curve. For almost all countries a steady
increase in the number of fatalities, followed by a steady decrease is found for the
umber of n
b
1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1250
1500
1750
2000
2250
2500
Fatalities: Greece

Figure 5.7. Observed and predicted fatality numbers for Greece, using vehicle fle
gures and fatalities in Harvey's structural time series model.
et
fi


In general, it can be observed in all three graphs that changes in the number of
fatalities from year to year are almost always within the error bounds. This means
that year-to-year changes which are sometimes considerable, as is the case for
69
Greece in the mid-1980s, are either explained by sudden changes in exposure, or
by mere chance.
5.4. Trends of fatality rates for grouped countries
Using the final grouping of countries from Section 4.4, a series of SVD analyses
were carried out on the data of the four groups. It turned out that for all groups the
first dimension was again the most important. The eigenvalues and the percentages
of explained variance by the first dimension are given in Table 5.1. For group 3 the
analysis in two dimensions gives a perfect fit, because only two countries were
used.

Eigenvalues

Dimension 1 Dimension 2 Dimension 3
% Explained
variance by
Group 1 5.426 0.561 0.248 98.58
Group 2 9.602 0.729 0.545 98.96
Group 3 4.050 0.427 98.90
Group 4 10.772 0.971 0.422 99.04
Table 5.1. Eigenvalues and the percentage of variance explained by the first
dimension from four SVD analyses of fatality rates in four groups of countries.


The trends of the first dimension for the four groups are given in Figure 5.8. The
trends are normalized such that the sum of squares of all values is 1. Because
group 3 only has rates from 1981 onward, its graph shows higher average values
than the graphs of the other groups.

The trends for group 1 and 2 are rather similar. The trends for group 3 and 4 have
peak values around 1990.

General trend in fatality rates in 4 groups of countries
0,4
0
0,05
1970 1975 1980 1985 1990
0,1
0,35
1995 2000 2005
0,3
0,15
0,2
0,25
T
r
e
n
d

f
a
c
t
o
r
Group 1 Group 2 Group 3 Group 4

Figure 5.8. General trend (first dimension) in fatality rates, resulting from SVD
analysis of four groups of countries.
70
The factor scores for the countries are given in Table 5.2. These factor scores can
only be compared within groups. A comparison of these factor scores for all
ountries was given in the second column of Table 4.2. c

Group 1 Group 2 Group 3 Group 4
Country
Factor
score
Country
Factor
score
Country
Factor
score
Country
Factor
score
Czech Republic
Hungary
0.458
0.889
Greece
Portugal
Spain
0.644
0.673
0.363
Denmark
Finland
Great Britain
The Netherlands
0.441
0.543
0.317
0.519
Austria
Belgium
France
0.467
0.368
0.372
Germany
Ireland
0.301
0.357 Norway 0.257
Sweden 0.274 Italy
Luxembourg
Switzerland
0.261
0.387
0.269
Table 5.2. Factor scores for the countries per group for the first dimension.
In the first group (Figure 5.9a), the Netherlands and Finland show a better improve-
behind


The residuals from the first dimension for each country in each group are given in
Figure 5.9a-d.

ment from 1980 onward than the other countries. Especially Denmark lags
during this period.

Residuals from trend in fatality rates for group 1
-0,3
-0,2
-0,1
0
0,1
0,2
0,3
R
e
s
i
d
u
a
l
s
1970 1975 1980 1985 1990 1995 2000 2005
Denmark Finland Great Britain Netherlands Norway Sweden

Figure 5.9a. Residuals from the first dimension of an SVD analysis of the fatality
rates of country group 1.


In the second group (Figure 5.9b) there seem to be two subgroups, with respectively
higher and lower values during the 1980s and 1990s: higher values for Ireland,
Belgium and France; lower values for Austria, Luxembourg, Switzerland, and
Germany; Italy's value is between those of the subgroups. The second subgroup
performs better than the first subgroup during this period.
71
Residuals from trend in fatality rates for group 2
-0,3
-0,2
0,0
0,3
1970 1975 1980 1985 1990 1995 2000 2005
s
i
d
u
a
l
s
0,1
0,2
-0,1
R
e
Austria Belgium France Germany Ireland Italy Luxembourg Switzerland

Figure 5.9b. Residuals from the first dimension of an SVD analysis of the fatality
rates of country group 2.


Figure 5.9c shows that the Czech Republic performs better before the early 1990s,
and Hungary after that period. In recent years these countries have been performing
almost equally.

Residuals from trend in fatality rates for group 3
-0,3
-0,2
-0,1
1980 1985 1990 1995 2000 2005
R
0
0,2
0,3
e
s
i
d
u
a
0,1
l
s
Czech Republic Hungary

Figure 5.9c. Residuals from the first dimension of an SVD analysis of the fatality
rates of country group 3.


Figure 5.9d shows that the increase in the fatality rate around 1990 is caused by
Spain and Portugal; the residuals for Greece are highly negative during that period.
rom 1994 onward the fatality rate developments are more similar for the three F
countries.

72
Residuals from trend in fatality rates for group 4
-0,3
-0,2
-0,1
0
0,1
0,2
0,3
1970 1975 1980 1985 1990 1995 2000 2005
R
e
s
i
d
u
a
l
s
Greece Portugal Spain

Figure 5.9d. Residuals from the first dimension of an SVD analysis of the fatality
rates of country group 4.


Finally, an SVD analysis was carried out on the fatality rates of group 1 and group 2
together. For this analysis the numbers of fatalities for all countries in group 1 were
he
irst
is 99.87%. Figure 5.10 shows this general trend, together with an
xponential best fit.
added together and divided by the total number of vehicles for these countries. T
same was done for group 2. The percentage of explained variance for the f
imension d
e

General trend in fatality rates in groups 1 and 2
0,4
0,3
0
0,1
1970 1975 1980 1985 1990 1995 2000 2005
c
t
o
r

0,2
T
r
e
n
d

f
a
Figure 5.10. General trend (first dimension) in fatality rates, resulting from SVD
analysis of group 1 and group 2, together with the best fitting exponential trend.


73
Residuals from trend in fatality rates in groups 1 and 2
-0,06
-0,04
-0,02
0
0,02
0,04
0,06
1970 1975 1980 1985 1990 1995 2000 2005
R
e
s
i
d
u
a
l
s
Group 1 Group 2

Figure 5.11. Residuals from an SVD analysis of the fatality rates in group 1 and
group 2.


Figure 5.11 gives the residuals for both groups. These values, varying from 0.04 to -
0.04, are rather small, but systematic.

In the beginning of the 1970s group 1 performs better than group 2. From the late
1970s until the 1990s a reversed trend can be noticed, which shows that group 2
erforms better than group 1. This is the same period for which the dimension
with the picture in Figure 5.1, which shows the developments in fatality rate
nd mortality rate. There smoother curves for all countries in the direction of the 0-0
roups of countries
disaggregate developments are given in Appendix 8.
In order to make comparisons easier, all data are summed over the years in five-
ars
These results are pictured in graphs (Figure 5.12a-d and Appendix 8). The group
p
values deviate most from the exponential fit which is given in Figure 5.10. The better
performance of group 2 is still present in recent years, although to a lesser extent.

If we look at the four Figure 5.9a-d, we see that in all groups the safety
developments tend to become more and more similar in the later years. It is as if all
European countries aspire to a common safety policy and position. This tendency
agrees
a
point can also be observed in recent years.
5.5. Analysis of disaggregate data for g
Using the selected group structure, three types of disaggregate developments are
compared for the countries within groups. The disaggregate data concern age,
traffic mode and road type. Some developments will be described in detail in this
section. All

year blocks, from 1970-1974, .., to 2000-2004. The last block consists of the ye
005-2007, as far as data is available for these years. 2

Age is classified in four categories: 0-14 years, 15-24 years, 25-64 years and 65+
years. For each country, the fatalities in each age group are summed over the years
in each 5-year block and divided by the total number of fatalities in that 5-year block.
74
average is also added to each graph. This way each country can be compared with
every other country in the group and with the group average.

Traffic mode is classified in three categories: 4-wheel motorized, 2-wheel motorized
and non-motorized. Again, the share of fatalities for each transport mode related to
the total number of fatalities is computed and the results and the group average are
added to the graphs of each group (Figure 5.13a-d and Appendix 8).
Road type is classified in three categories: urban roads, country roads and
motorways. The same procedure is used for this disaggregate data (Figure 5.14a-d
and Appendix 8).

For each group, each disaggregation, and each category this data is plotted,
resulting in 4 x 4 + 4 x 3 + 4 x 3 = 40 graphs in Appendix 8. Only one category for
each group and one disaggregation will be presented in this section.

For group 1 (Figure 5.12a) one might say that the proportion of children among road
fatalities decreased from around 11 percent in the early 1970s to around 3 percent
around 2005. Overall, Sweden has the lowest share of children in road fatalities;
Norway has the largest decrease, from around 16 percent to 2 percent.


Group 1: 0-14 years
0,04
0,1
0,12
0,16
0,18
P
r
o
o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0,14
0,06
0,08
p
o
r
t
i
o
n

0
0,02
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
Denmark Finland Great Britain Netherlands Norway Sweden Average

Figure 5.12a. The proportion of 0-14 year old children in the road fatalities of
group 1.


Figure 5.12b shows that the proportions in group 2 are similar for most countries
except for Ireland and Luxembourg. The proportion of children in road fatalities is
approximately 9 percent in the early 1970s; this percentage is lower than for
group 1. The share of children in road fatalities decreased to around 3 percent after
2000, a percentage that is comparable to that in group 1. The drop is most evident
for Ireland: a fall from around 12 percent in 1980 to almost 2 percent in 2005. The
data for (the small country) Luxembourg is a bit unstable. Since the 1980s, Italy has
by far the smallest proportion.

75
Group 2: 0-14 years
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland
Italy Luxembourg Switzerland Average

Figure 5.12b. The proportion of 0-14 year old children in the road fatalities of
group 2.

Groups 3 and 4 (Figure 5.12c and d) show the same general picture: a drop over
time to a final 3 percent in recent years for group 3 and two percent for group 4.

Group 3: 0-14 years
0,02
0,03
0,04
0,05
0,06
0,07
0,08
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0
0,01
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
Czech Republic Hungary Poland Slovenia Average

Figure 5.12c. The proportion of 0-14 year old children in the road fatalities of
group 3.


76
Group 4: 0-14 years
0
0,02
0,04
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f
0,06
0,08
0,1
r
o
a
d

f
a
t
a
l
i
t
i
e
s
0,12
Greece Portugal Spain Average

Figure 5.12d. The proportion of 0-14 year old children in the road fatalities of
group 4.


If the outcomes of the four groups are compared, the proportion of fatalities for
young children can be seen to decrease with time for all groups. Initially, the
variation between groups and within groups is larger than in recent years. For
groups 1, 2 and 3 the proportion is around 3 percent in recent years. The southern
countries (Italy included) have the lowest percentage of fatalities among children
between 0-14 years old. In those countries this percentage has been just above 2
percent in recent years.

Figure 5.13a-d show the proportions of non-motorized road users in fatal crashes.
From group 1 (Figure 5.13a) it follows that all countries show the same decreasing
trend for the proportion of non-motorized road fatalities as for that of the children.
The percentages decrease from around 35 percent to 25 percent. Of course this is
partly due to overlap: many of the non-motorized road users are children. Sweden,
Norway and Iceland do better in this respect than the other countries in group 1.

For group 2 (Figure 5.13b) the decrease is a bit higher, from around 35 to 20
percent. Once more, the proportion is highest for Ireland, but from the 1990s onward
the decrease continues to reach the group average in approximately 2005. The
percentage is lowest for France.

77
Group 1: non-motorized
0,1
0,2
0,3
0,4
0,5
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
Denmark Finland Great Britain Iceland
Netherlands Norway Sweden Average

Figure 5.13a. The proportion of non-motorized road users in the road fatalities of
group 1.


Group 2: non-motorized
0,6
0,4
f
a
t
a
0
0,1
0,2
0,3
P
r
o
p
o
r
t
i
o
n

o
f
0,5

r
o
a
d

l
i
t
i
e
s
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
Austria Belgium France Germany Ireland Italy Switzerland Average

Figure 5.13b. The proportion of non-motorized road users in the road fatalities of
group 2.


The situation is different for group 3 (Figure 5.13c). Here the drop is from around 50
percent to 40 percent non-motorized victims. Only Slovenia is doing much better,
nd its performance can be compared w a ith group 2.

Group 4 (Figure 5.13d) is also comparable to group 2, with a drop from 35 percent
to 20 percent.

78
Group 3: non-motorized
0,2
0,3
0,4
0,5
0,6
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0,1
P
r
o
p
0
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
Czech Republic Hungary Poland Slovenia Average

Figure 5.13c. The proportion of non-motorized road users in the road fatalities of
group 3.


Group 4: non-motorized
0,3
0,4
0,5

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0
0,1
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n
0,2
Greece Portugal Spain Average

Figure 5.13d. The proportion of non-motorized road users in the road fatalities of
roup 4. g


All groups again show a tendency for the percentage of non-motorized victims to
decrease over time. In contrast with the previous decrease in the share of children,
however, initial differences between countries within groups seem to remain
unchanged over time. This results in larger differences between countries within
groups, as well as between groups in recent years.

It should be realized that the observed drops in share of non-motorized victims (and
the previous ones for children) are relative, compared to the fatalities for motorized
road users (or other age groups). However, given the steady overall decrease in the
number of fatalities over the years it can still be said that the drops in the number of
fatalities for non-motorized road users and children are the most significant. Road
79
use has become much safer for the most vulnerable road users during the last 35
years.

Figure 5.14a-d show the proportions of fatalities on urban roads. Figure 5.14a
shows a general tendency for this proportion to decrease over time. Of group 1,
Great Britain has the highest overall percentage, while Finland and Sweden have
e smallest proportion of fatalities on urban roads. th

Notwithstanding the overlap with the group of children and the non-motorized road
users, the drop from around 40 percent to around 30 percent is smaller and more
diverse for the fatalities on urban roads than for the young victims in this group of
countries.

Group 1: urban roads
0,3
0,4
0,5
0,6

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0,2
o
p
o
r
t
i
o
n
0,1
P
0
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
r
Denmark Finland Great Britain Netherlands Sweden Average

Figure 5.14a. The proportion of fatalities on urban roads in countries of group 1.


The same pattern is seen for group 2 (Figure 5.14b). Here the drop is also from
round 40 percent to around 30 percent and the a variation between countries is
much larger than was the case for the proportion of 0-14 year old victims. All
countries show this decrease in share of fatalities on urban roads, except Italy.

For group 3 (Figure 5.14c) the share of fatalities on urban roads drop from around
55 percent to 43 percent. These percentages are considerably higher than for
groups 1 and 2. The proportions for Slovenia are again considerably lower than
those for the other countries in group 3.

80
Group 2: urban roads
0,5
0
0,2
0,4
0-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
o
r
t
i
o
n

o

f
a
t
a
l
i
t
i
e
s
0,3
f

r
o
a
d
0,1
P
r
o
p
1970-1974 1975-1979 198
Austria Belgium France Germany Ireland Italy Switzerland Average

Figure 5.14b. The proportion of fatalities on urban roads in countries of group 2.


Group 3: urban roads
0,5
0,6
0,2
0,3
0,4
p
o
r
t
i
o
n

o
f

r
o
a
d
0
0,1
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average

Figure 5.14c. The proportion of fatalities on urban roads in countries of group 3.


For group 4 (Figure 5.14d) only data from Portugal and Spain is available. Here we
percentages are very different but stable over
time. For Portugal the percentage is around 45% and for Spain around 18%.
see a completely different trend. The
Because there is a major overlap of fatalities on urban roads with age and travel
mode, this suggests that the relative safety on urban roads has deteriorated since
the 1980s. The same conclusion was already found for Italy, where the proportion
even increased over time. In this respect the southern countries differ strongly from
the other groups.
81
Group 4: urban roads
0,3
0,4
0,5

r
o
a
d

f
a
t
a
l
i
t
i
e
s
0
0,1
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
0,2
p
o
r
t
i
o
n

o
f
Portugal Spain Average

Figure 5.14d. The proportion of fatalities on urban roads in countries of group 4.
5.6. Conclusions and recommendations
Inspecting all these graphs and more generally all the results of the analyses
iscussed in this chapter, we may conclude that d all European countries tend
towards the same aggregated or macroscopic level of road safety (combination of
personal safety and traffic safety) however the speed of improvements differ. When
considering their focus on avoiding special types of accidents, there are important
differences between the individual countries as well as between groups of similar
countries in terms of how they reach this level of road safety. In other words, the
general policy of improving road safety in each country ultimately could result in the
same level of safety as that in other counties, but that level is achieved in different
ways for different countries tackling different types of problems. More detailed
analyses, such as those carried out in earlier SUNflower studies, are needed to
understand these differences. It is recommended to carry out these analyses using
methodologies such as SVD analysis and structural time series analysis.


82
6. Application of SUNflower at regional or city level
SUNflower analyses to da
lessons that can be learne
te have focussed on national comparisons and the
d by individual countries when comparing their road
level where an authority can
ave been
e, having something to say about road safety. It can be claimed, that
country.
safety policies and resulting performance with that of other countries. A similar
scope exists for the comparison of programmes and their performance at the sub-
national level which consists of regions and municipalities.

n principle, such comparisons could be made at any I
pursue its own safety programme and policies or where developments in
motorization or in economic activity result in changes in traffic levels by different
modes, leading to changes in safety outcomes. Given the fact that policies were
applied centrally by national authorities, it is straightforward to expect such
comparisons to be most fruitful at national level. But in general, a comparison of
performance is likely to be rewarding where the authority has sufficient autonomy to
significantly influence the progress of safety performance in its area, and where the
area is large enough to yield sufficient accident data to enable robust comparison of
output measures. This is likely to be the case for either regions or large individual
cities.

Historically, in Europe cities and regions have played an essential role in the
formation of national-states and national societies. European integration has
contributed to a certain relaxation of the 'constraints of state' and cities and regions
have emerged as one possible level for regulation of interests, groups and
stitutions. Large national supervisory organizations and institutions h in
losing their dominance in the context in which there is no longer one single centre
where governance is concentrated and the sub-national levels should be considered
to be of the same order as those of states or of the European Union. We have
already been witnessing a movement in the redistribution of political power away
from Member states to the advantage of sub-national governments (Hooghe, 1996;
Le Gals, 1998). European structural funds in particular, have contributed to the
development of the European polycentric governance. In parallel, the 'new
governance' that is emerging is concerned with 'soft' forms of policy-making based
on voluntary commitment, coordination in place of authoritative instruction, and
subordination to common goals. In this context, new actors (stake-holders) appear
n the stag o
territories have become the trial fields of numerous advances in road safety science,
such as ISA, Alcolock, or the shared-space urban road design concept. Cities and
regions should however not be seen only in the terms of governance, but also as
behaviour-shaping areas, with their own culture, norms and habits.

Although the sub-national road safety performance will be influenced in part by
national policies (just as national polices are influenced by international vehicle
regulations), comparison at the lower level could be particularly useful where there
are substantial differences in mobility patterns, social-economical conditions,
residential and industrial densities or road network patterns, or where specific
policies which are not evident on a national basis have been followed locally. The
territorial level pertinent for such analysis should be given by the governance
tructure of a particular s

Initial SUNflower comparisons of Sweden, the United Kingdom, and the Netherlands
reflected that these three countries had reached similar levels of safety performance
83
(in terms of overall national statistics) despite the differences in population density
and road networks. It was of interest therefore to see how this had been achieved.
Each country was faced with similar general problems in terms of risk in driving
behaviour speeding, drinking and driving, seatbelt wearing - and their performance
in these areas could therefore be compared directly. Lessons learned might be
directly transferable. The observed differences in modal split reflected both cultural
es in population density,
accessibility, social deprivation, but also investment in road networks. Each country
terms of knowledge and road safety
improvements which can be achieved through an approach which has so far been
djustment of the method which has so far been applied at the
national level. The purpose of this chapter is to explore the potential for comparisons
of safety performance at th onal le se methods allowing
reliable comparison.
6.1. Differences in application at th ional level
The differences between individual regions ar d out when national accident
statistics are quoted. When comparing road safety outcomes at a regional level or
b w, ary for an l analysi safety outcomes to take into
account substantial structur ces such as demography, urbanization, social
fic
s, for example
ural environments, and a similar
nge of policies. To understand more fully the factors driving differences between
safety performance it is therefore important to seek areas with differences in mobility
patterns, modal split, network configurations, or in safety programmes or in attitudes
towards safety improvement, either culturally or organizationally.

and spatial differences, which in turn partly reflect differenc
was faced therefore with specific issues in tackling its own modal safety problems
e.g. cyclists and moped riders in the Netherlands, pedestrians and motorcyclists in
the UK. The size of these specific safety problems and the extent to which each
country had been effective in tackling their own major accident contributors would be
reflected in the national casualty rates.

Thus there is a potential for innovation in
left unexploited: the application of the SUNflower method at sub-national level. The
potential beneficiaries are not only local administrations, in preparing and applying
programmes and policies at local level, but also national administrations in better
understanding of structural factors influencing the application and results of road
safety policies. The application at the sub-national level might however require a
rethink and an a
e sub-nati vel and to propo
e sub-nat
e average
elo it is necess y rationa s of road
al differen
deprivation, which have a direct impact on accident risk factors, such as traf
olume, flow composition, traffic speed, and road network layout. v

Some account of these factors can be taken at the national level by analysing using
rural and urban statistics, or by looking at road user groups separately where those
groups (e.g. pedestrians) are likely to be primarily at risk in urban environments.
Similarly, analysis by separate road groups allows some difference between types of
area to be shown. But such analyses, for example, still average all data from urban
areas together rather than enabling differences between different urban areas to be
xplored. Also the traffic performance of some transport mode e
powered two-wheelers, differs between urban and rural areas, which results in
different modal patterns in these areas.

Similar problems arise when regions are compared which mainly reflect a sub-set of
ational data, with similar proportions of urban and r n
ra
84
At the sub-national level there are a number of methodological issues which need to
be considered compared to national approach. First, there are significant differences
in the definitions of regions, which may have very different properties. For example,
there are different ways of defining urban regions: administrative approach
(legal/administrative status), morphological status (extent/continuity of urban space)
and functional approach (core area vs. surrounding territory). So, clearly, the data
compatibility remains an important issue. Second, analyses of aggregated data
bring along potential errors in the interpretation of statistical data, such as Modifiable
Area Unit Problem (MAUP), or ecological fallacy. The studies which compare lower
administration areas are more likely to suffer from MAUP and ecological fallacy.
6.1.1. Potential uses of sub-national comparison
There are two basic reasons for comparing the safety performance of sub-national
areas.

The first is to provide a ranking of the relative performance of each area this will
way,
milar
within the ranking process. This could
e done by developing a model of safety outcomes which includes these factors.
e defined as individual relatively large urban areas; the
be most useful for comparison within countries. To do this in a meaningful
factors which affect the ability of safety practitioners to achieve si
provements in safety need to be included im
b

The second is to provide better understanding of the factors affecting safety
improvement, so that safety practitioners can achieve more effective programmes.
This requires greater focus on understanding how the effects of programmes are
modified by the nature of the safety problem faced by each area. Lessons can be
learned not only from comparison within countries, but also from comparison
between similar areas in different countries. This might be more effectively achieved
by case studies comparing small groups of areas, than by attempts to model a wide
range of areas.

The areas of most interest will be those taking prime responsibility for spending local
udgets on safety programmes. In some countries, these will be regions, in others b
they will be individual urban areas, in yet others it will be a combination of both. It is
likely that a modelling approach will be more appropriate for comparing regions,
while city authorities might find a case study approach more effective. But the
appropriate analysis setting should be determined individually for each Member
state, rather than through a blind general application of corresponding administrative
boundaries. The two types of territory will be discussed separately later in this
chapter.

Regions are regarded in this context as parts of a country, usually administratively
defined spatial units, which combined make up the national territory; they may be of
varying sizes and may be largely urban, largely rural, or a combination of urban and
ural areas. Cities ar r
boundaries of a city area are likely to be those for which relevant statistical data are
available because it also has a particular administrative or legal status.
6.1.2. Understanding factors affecting sub-national comparison
There are two ways to make any statistical sub-national comparison meaningful and
reliable. One is to compare areas that are similar in terms of their physical structure
such as road network configurations, climate, relief, mobility patterns, modal split,
85
and attitudes towards road safety. The second is to try to adjust for as many of
these factors as possible. The former way limits the number of areas available for
comparison, the latter is sensitive to data reliability and availability, and implies the
use of sophisticated statistical methods which could make the comparison less
understandable to broader public and bring about the risk of the use of inappropriate
rison made for a longer period of time,
n the other hand, would require considering the changes in structural factors
behind accident risk, such as demographical or socio-economical development.
lore
ent or absent in situations where good or poor safety
erformance appears to have been achieved. This approach allows factors that may
at sub-national level than at national
lly not at stake, the problem of scarcity of
wer approach uses three road safety indicators:
and confounding factors.

At the same time, no barriers exist to carry out a meaningful and reliable comparison
of time trends within countries over a shorter period of time, although this may also
require the use of more complex statistical models due to scarcity and fluctuation of
analysed road accident outcomes. Any compa
o

An alternative approach, consistent with the SUNflower case studies, is to exp
he factors that are pres t
p
be difficult to define in numerical terms, such as organizational efficiency or the
presence of constraints on the implementation of safety policy to be considered
alongside numerical characteristics of an area. The outcome of such an approach
will be less quantitative but allows a potentially wider range of factors to be
onsidered. c
6.1.3. Applying the pyramid approach
6.1.3.1. Safety outcomes
Accident statistics are likely to be less robust
level. While the reliability is genera
accident outcomes becomes a serious problem raising questions about the
significance of any differences that are identified. There are two ways of overcoming
these difficulties. First, it is possible to work with the accident outcomes of higher
frequency (killed and serious injuries, or all injuries) and the second way is to turn to
the advanced statistical methods that are commonly used in epidemiological
research (e.g. Bayesian spatio-temporal models).

The matter of using an appropriate measure of safety becomes even more important
t the sub-national level. The SUNflo a
i.e. fatalities per head of population, fatalities per licensed vehicle, and fatalities per
vehicle kilometre. Safety performance within a region will be strongly influenced by
the amount of vehicular travel, and thus fatalities (or injuries) per vehicle kilometre
(fatality or injury risk) is likely to be the most relevant measure. While amounts of
travel within a city will also be reflected in safety performance, modal split and
interaction between modes become more important factors in these areas and may
be included among the explanatory factors.

The most direct measure of the amount of vehicular travel in an area is the length of
roads of different types multiplied by the Average Annual Daily Traffic (AADT) on
these roads. In the absence of such data, an estimate of vehicle kilometres may be
made by developing a relationship between the amount of travel per inhabitant (in
terms of billion vehicle kilometres) and the population density. The relationship
between the population density and the travel per head of population is typically
exponential, which is illustrated for French regions in Figure 6.1, showing higher
86
traffic performance per r
re
esident population in less densely populated regions. This
flects longer commuting distances for rural area residents, greater availability of
public transport services in urban areas at individual level, and more transit traffic
and less commercial trips at aggregated level in rural areas.

Traffic performance at different population density
0
100
200
300
400
500
600
700
800
900
1000
0 2 4 6 8 10 12 14 16
P
o
p
u
l
a
t
i
o
n

d
e
n
s
i
t
y
Vehicle kilometres per head of population (in 1000)

Figure 6.1. Population density against motorized vehicle kilometres per head of
population for 22 French regions in 2004.


It can be concluded, that the best measure of safety at the regional level would be
the one taking account of the real exposure of road users in road traffic, typically in
terms of vehicular travel.
.1.3.2. Perf 6 ormance indicators
gions (Taggi et al., 2006).
onal sample studies. Local road safety
erformance indicators may suffer from sampling errors, due to the limited size of
Most European countries can only set road safety performance indicators at a
national level. Detailed analysis suggests that even between a country's regions
there are important differences in cultural, linguistic and economical background.
For example, there are significant differences in the use of restraint systems
etween Swiss (Siegrist et al., 2006), or Italian re b
Similarly, the prevalence of driving while intoxicated is expected to vary between for
example the French, Italian and Spanish regions. As for speeding, the detailed data
which allow determining the real extent of these disparities unfortunately are rarely
vailable. a

More detailed observation also shows that certain performance indicators vary for
different road types or for different types of urbanization. For example, vehicle
occupants are more likely to use their seatbelts on longer trips on motorways and
main roads than in cities. These nuances should be accounted for when considering
small administrative units, particularly cities.

Although regional or local data on casualties will generally be available, they may
not allow for analysis of causal factors to the same extent as at national level and
herefore it may be necessary to rely on nati t
p
87
the samples used for setting sub-national indicators. Similarly, some behavioural
surveys may be conducted locally, but unless local differences are anticipated and
have been assessed, it may also be necessary in this case to assume that
performance indicators reflect national sample surveys.
6.1.3.3. Measures and programmes
The polycentrism in road safety management is apparent in many countries which
have established local road safety targets, strategies and programmes. The extent
f the responsibilities and spending power of regional and local governments are the
e are also significant differences in the
application of concrete measures and policies.
Data on individual measures and implemented programmes can be expected to give
he report on the first SUNflower project (Koornstra et al., 2002) recorded the
tion density
imilar to the Netherlands. The average annual traffic flow in Britain is higher than in
e problems in the Netherlands and in Britain, respectively, when
these modes of transport are also taken into account. If these increased risks do not
occur, specific road safety measures for the protection of these particular road user
ese countries.
o
most determinative factors of their potential impact. But it is not only at this strategic
and administrative level of governance that the potential differences in road safety
outcomes could be conceived, as ther

a clearer picture of the safety outcome at the local level than when the data on the
large number of individual programmes that make up measures of national activity is
averaged. Differences between local areas that have implemented particular
measures and those that have not can also be explored more clearly. However, the
same as at the national level, information on the sub-national level of
implementation is likely to be of variable quality. Analysis at city level will enable
engineering measures to be related more closely to particular types of roads.
Information on the expected effectiveness of different measures will still need to be
based on wider studies as these will be based on larger numbers of accidents.
6.1.3.4. Structural (spatial and demographic) differences
T
differences in population density, road density, and percentage of vehicle kilometres
on motorways in Sweden, the United Kingdom and the Netherlands. Population
density, road density and motorway traffic percentage were highest in the
Netherlands and lowest in Sweden. Within the UK, England has a popula
s
the Netherlands, and the average traffic flows in Sweden are only a third or quarter
of that in the other two countries. The report states that the car occupant risk could
therefore be expected to be highest in Sweden. Cyclist and pedestrian safety could
be considered larg
groups have probably been taken in th

Koornstra et al. (2002) showed that although the overall fatality rates in the three
countries were similar, the rates for individual transport modes differed, with the
fatality rate for car occupants being highest in Sweden, while the rates for
pedestrians and cyclists were highest in Great Britain. This was largely as expected,
except that despite the large cycle flows cyclist fatality rates were low in the
Netherlands. The large flows induced many risk reducing countermeasures, for
example cycle facilities such as cycle tracks. The report of the extensive analysis of
the three original SUNflower countries (Lynam et al., 2005) includes an estimate of
the likely effects of traffic flow on the national accident rates for individual transport
modes and road types. Quantifying these effects is not easy, but based on average
activity rates for the different transport modes in each country; the report suggests
88
that they might explain up to half the difference in car occupant fatality rates and
most of the difference in pedestrian fatality rates. But the report concludes that much
ore detailed data and analyses are required to explore these effects properly.
es. Analysis at sub-
ational level, therefore, offers the potential to make real advances in understanding
et of consistent values and measures
together form the value system, which is subjective and varies across populations.
Types of value include ethical/moral values, ideological, social and aesthetic values,
l attitudes of road
e
ounds. The
rentino-Alto Adige/Sdtirol region, for example, was part of the Austria-Hungary
m

Lynam et al. (2005) also considered the role of town size and of the presence of
economically disadvantaged groups in explaining differences in pedestrian fatality
rates between the three SUN countries. There are indications that both these factors
have an effect on pedestrian fatality rates, although the role of economically
disadvantaged groups mainly appeared to be important in Britain (Broughton &
Buckle, 2005). However, the role of these factors was not quantified in any detail,
and for comparison at the sub-national level, a much more in-depth analysis would
be justified.

It is likely that data collected locally on these issues will be more relevant to safety
outcomes than approximation of this data to national outcom
n
the role of policy, process and organizational structure in safety outcomes. Such
data may require questionnaire surveys, and in some cases in-depth interviews with
relevant regional or municipal officials to more fully understand the decision-making
processes, and the approaches used to gain acceptance of policies.
6.1.3.5. Cultural differences
In previous descriptions of the safety pyramid, structural and cultural issues have
been grouped together in the lowest level of the pyramid. Both are important to
understanding the subsequent levels, but they act in different ways, and it can be
argued that they might be defined as separate levels.

Many regions and towns were built upon a common cultural background. This could
potentially influence road user behaviour and, consequently, the way in which road
safety is managed at the sub-national level. Here, culture is regarded as reflecting
values and norms in their social sense. A s
and it may be argued that all of them have an effect on behavioura
users, which in turn manifests itself in different road safety outcomes. Concret
values such as the value of a human life, respect for each others rights, etc., are
directly reflected in road safety provisions, such as those related to casualty
reduction targets. Norms refer to the rules that are socially enforced. Social
sanctioning is what distinguishes them from values. They can be viewed as
reference standards, or statements that regulate behaviour and act as informal
social controls. The most typical example is society's drink-driving attitude, which
differs significantly among countries, but may also vary between regions.

It can be argued that national borders only partly reflect cultural differences between
European citizens. Conflicts and political ideologies have produced some national
border adjustments which do not reflect historical and cultural backgr
T
Empire until its annexation by Italy in 1919 and German speakers still represent
more than a third of the population. Another obvious example of cultural variation is
Switzerland with its three language communities. Hence national cultural
characteristics cannot be expected to be common to all regions within a country.
89
However, these differences seem to have a rather limited effect on road safety
oposed division between 'structure'
and 'culture' is the structure of the organizations that are to develop and implement
p re can be regarded as the result of management
actions, and as such becomes part of the development of effective safety policies
articular attention when applying the SUNflower
pproach at a regional level. The substantial effect of physical structure on road
2. Role of (physical) structural differences
veloped in the ESPON
3
project,
and one of the indicators was settlement structure. Settlement structure typology

outcomes.

One aspect which potentially transcends the pr
safety olicy. Organizational structu
and programmes. But organizational structure at the sub-national level will inevitably
be affected by national governmental structures that will be influenced by national
cultural characteristics. In comparison with middle and south European countries, for
example, Anglo-Saxon countries have a long tradition in accountability. This leads to
a broader application of results and performance auditing, potentially influencing the
effectiveness of policies. The extent to which effective organizations are present at
the sub-national level will also be depend on the values espoused by local safety
champions.

Of all the pyramidal levels, those addressing structure in both physical and
organizational sense deserve p
a
safety at sub-national level is discussed in Section 6.2 below. Knowledge on the
mechanisms and impacts of organizational features of road safety management is
much more limited, but this is briefly discussed in Section 6.3.
6.
A more detailed analysis of road safety outcomes at sub-national level unveils the
existence of significant differences in risk indicators between regions. They are of
the same order as the differences in risk between European countries (Eksler et al.,
2008a). Given the fact that the number of factors influencing road safety outcomes
at a sub-national level is limited compared to those at the national level, it is
straightforward to assume that structural differences between regions may explain
part of the variation in road safety outcomes.
The classification indicators for EU regions were de
was established by Schmidt-Seiwert (1997) based on the two criteria population
density and size of centres. They distinguish agglomerations, urbanized areas, and
rural areas and these area types are used to define the following six types of regions
(see Table 6.1).

3
ESPON European Spatial Planning Observation Network, www.espon.eu
90
Region type Centre size Population density
1 > 300/km
2
Agglomerated region >300,000
2 150-300/km
2
3
150-300/km
2

[or <150/km
2
with a bigger centre >300,000]
4
Urbanized region 150,000-300,000
<150/km
2
5 > 125,000

6
Rural region
< 125,000
<100/km
2

Table 6.1. Settlement structure typology versus population and density (ESPON).


A generalized linear model (GLM) was run, assuming road mortality to be a function
of settlement type and country (to account for different levels of road safety in
4 for altogether 250 NUTS
4
-2 regions of 25
nalysis (Eksler, to be published). NUTS-2
3.0 million and may correspond to
therlands, Belgium).
different countries). Data of the year 200
EU countries were considered in this a
regions have typically a population of 0.8 to
regions (e.g. France, Italy) or provinces (the Ne

Except the one for densely populated areas with large centres, the regression
coefficients for particular settlement types appear to be significant. The average
road mortality ratios for the six different settlement types in relation to national and
European averages are summarized in Figure 6.2.

Mortality rate for different settlement types
0
0,5
1
1,5
2
2,5
R
e
l
a
t
i
v
e

m
o
r
t
a
l
i
t
y

r
a
t
e
Type 1 Type 2 Type 3 Type 4 Type 5 Type 6
Region type

Figure 6.2. Relative average mortality rate for different settlement types of 250
NUTS-2 regions of 25 EU countries in 2004.


Clearly, the settlement structure has a strong influence on the registered road
mortality at the regional level. Very densely populated regions with large centres
have an average mortality rate which is less than half that of the average, while less

4
NUTS - Nomenclature of Territorial Units for Statistics (EUROSTAT). The NUTS classification is
based on population criteria, with NUTS-1 regions having typically 3 to 7 million inhabitants, NUTS-2
regions 0.8 to 3.0 million, and NUTS-3 regions 0.15-0.8 million.
91
densely populated regions without centres, which are rural areas, have a mortality
rate that is up to twice as high as average. Relatively low mortality rates in urban
areas could be partly explained by larger public transport use and shorter trips which
reduce exposure to risk on the one hand, and, on the other hand, by low travelling
speeds due to congestion resulting in less serious injuries. High mortality rates in
rural regions probably reflect high travelling speeds, often on undivided roads, and
longer trips. Using vehicle kilometres instead of population as the measure of
exposure significantly reduces the variance in risk between different types of region.

Since the settlement classification which is introduced above is based exclusively on
population density and centre size, it may be useful to introduce a more
sophisticated typology which takes functionality and morphology of area units into
account. In Belgium, an urbanization classification established by Luyten & Van
Hecke (2007) is widely accepted. Municipalities are subdivided on the basis of
several functional and morphological variables such as population density, number
of houses, and the size and date of construction of dwellings. It results in the
following subdivision for each municipality: city centre, inner city area, inner suburbs,
outer suburbs and rural areas. This subdivision gives a better definition of urban
areas as regards habitat and road network structure than the classifications based
on only one variable.

A comprehensive analysis of fatality risks (fatalities per vehicle kilometre) in 589
Belgian municipalities over the period 2001-2006 (Eksler & Lassarre, 2008, and
Eksler et al., 2008b) shows that on average peri-urban municipalities register 35%
higher fatality risks than core cities. The complete results which are summarized in
Figure 6.3 are as follows: city centre and inner city areas have lower fatality risks (of
al rate of 22% is larger
deserve more attention from road safety
21% and 7% respectively) and the latter's decline at an annu
than the urban area average. In inner suburban areas, the fatality risk is falling 33%
faster than average, possibly as a result of the implementation of traffic calming
schemes and other measures. In outer suburban areas, on the other hand, the risk
is 15% higher compared with the average and is actually going up. Rural
municipalities, which are not subject to the influence of urban areas, have a 10%
higher fatality risk, and the average annual risk reduction is 5% lower than average.

This analysis provides a new view of the simplified urban-rural area typology and
hows that it is actually the suburbs which s
professionals rather than the rural areas. However, this may be a phenomenon that
is specific for Belgium, as not all countries have had a similar degree of
decentralization of manufacturing, commerce, retail and office work resulting in new
waves of suburbanization (Champion, 2001).

92
Fatality risk for different settlement types
0,6
0,8
1,0
1,2
1,4
t
a
l
i
t
y

r
i
s
k

/

r
e
d
u
c
t
i
o
n

r
a
t
e
0,0
0,2
0,4
City centre Inner city Inner suburbs Rural Outer suburbs
R
e
l
a
t
i
v
e

f
1,6
a
Risk Annual reduction

Figure 6.3. Relative fatality risks and their relative annual reduction for different
Belgian settlement types.


One weakness of this approach is the lack of harmonization in the classification of
municipalities according to their urban structure across EU countries. At the same
his simple analysis gives us a very first suggestion of the structural factors which
may have an indirect impact on road safety outcomes. The review of related
in any
.2.1.1. Explaining differences in risk
y is linked with a 3.2%
er showed that the population density explained between 6
time, this is a very promising approach, which allows classifying spatial units into
more or less uniform groups for which a more reliable comparison could be made.

T
literature gives a rough idea of the structural factors that are to be considered
regional comparison.
6
Eksler (2007) investigated the effect of the differing demographic structures of 25
European countries on their road mortality rates (fatalities per head of population).
He found that at the national level, demographic structure, defined as the distribution
of age and sex in aggregated age groups (as also used in IRTAD) had only a minor
effect on mortality ratios. There can be greater effect when comparisons are made
between regions. He further concluded that adjusting for the demographical
structure when analysing mortality ratios could account for up to 12% of the variation
crude mortality rates. in

Eksler et al. (2008a) have shown that population density is probably the single most
powerful explanatory factor behind road mortality in Europe. The logarithmic
relationship between road mortality and population density at the regional level is
distinct in the majority of EU countries. This is illustrated in the Figure 6.4 showing
the log-log relationship between both variables for 21 EU countries. For many
countries, the region containing the capital city is included alongside all the other
regions in the plot. Based on the data from 1,089 regions of 25 European countries,
ey concluded that a 10% increase in population densit th
decrease in road fatalities per head of population. The level of spatial disaggregation
is not crucial here, as the similar results appear at different aggregation level.
Additional analysis furth
93
and 98% of the variation in road mortality registered at the level of NUTS-3 regions
in different EU countries, with a 59% average. Taking account of differences in
erally quoted, using fatalities
er head of population.
population density between European countries would thus produce a different
ranking of countries in their mortality rate than that gen
p

Figure 6.4. Relationship between road mortality and population density (log-log
scale) for NUTS-3 regions of 21 EU countries in 2002.


Population density as a synthetic indicator stands for a wide range of explanatory
ctors influencing the occurrence of fa fatal injuries in road traffic. Analyses performed

Among additional gional variation in
accident outcomes, we may find household income, employment rate, weather, or
alcohol use.

Annual household income could also be considered as a possible explanatory
variable behind accident outcomes at sub-national level. The effect of annual
average income on fatality risk (elasticity) estimated in the model run for 589
municipalities of Belgium for the period 2001-2006 is negative and has the value of -
0.239, meaning that a 10% increase in annual income per person is associated with
a 2.4% reduction in fatality risk (Eksler et al., 2008b). One possible explanation for
this is that people with higher household income tend to live in areas less affected
at the level of single countries suggest that the exposure in terms of kilometres
driven by motorized vehicles is the principal partial factor behind the population
density accounting for up to half of its explanatory value. This is basically because
the average annual distance driven per inhabitant of rural areas is generally higher
than that of urban citizens. Accident severity factors such as road environment or
avelling speed account for approximately 20%. The remaining 30% can be related tr
to the demographic, deprivation, economic and other factors (Eksler, to be
published).

Similar to population density, the urbanization rate can be used as a proxy standing
for a number of structural factors. In fact, both variables are strongly correlated, and
this results in similar conclusions to those described earlier for the effect of
population density. The differences in the definition of urbanization, however,
present a constraint to its use as explanatory factor. re

socio-economic indicators claimed to explain re
94
by transit travel, use safer mode of transports and benefit from better infrastructure,
which may result in lower accident severity.

Several authors have suggested that there are distinct and substantial effects on
casualty rates from characteristics associated with area deprivation across diverse
environments. These associations exist over and above influences arising from local
environmental characteristics. Distinct dimensions of deprivation appear to affect the
incidence of pedestrian and child casualties to varying degrees and sometimes in
different directions. This is likely in part to be related to the higher exposure to traffic
of these groups, but is also influenced by social factors (Noland & Quddus, 2004;
Christie, 1995; Graham & Stephens, 2008). The role of deprivation has been widely
studied in the UK, where the concept is well established, but not in other countries.

Practically all structural factors discussed above are correlated and their
simultaneous use should be done with care; otherwise analyses will result in
confusion rather than in understanding.
he differences indicated by population density reflect a complex set of factors that
tigate their interdependency.
his could be attempted at an aggregate level - for example, by adding more
Lassarre (in Delorme & Lassarre, 2005) attempted to explain the gap in fatality risk
in terms of behavioural factors associated with
usation and accident seriousness factors. These
riving, and seatbelt usage, and structural
factors such as urbanization and traffic patterns. He concluded that three major risk
6.2.1.2. Scope for extended SUNflower analyses
T
are likely to be associated with this density. Examples are numbers and lengths of
trips, choice of mode, and road conditions (e.g. traffic flows) faced by each mode.
These in turn will also be affected by the quality and quantity of the networks
provided for each mode. To gain deeper understanding of the issue, and how it
might be included within SUNflower type analyses, it would be useful to explore the
direct effect of changes in these factors, and to inves
T
variables to the type of model already used by Eksler & Lassarre (2008). Alternative
models, such as that used in Harland (1999) which focussed on fatalities to
individual road user groups and used local car ownership data, might also be
considered.

More directly, an attempt might be made to predict travel kilometres, both vehicular
and pedestrian, that might arise as a result of population factors, income, and
transport provision, and then relate the expected numbers of fatalities to travel
kilometres taking into account network quality and road user behaviour.

There is also further potential scope for extending the previous SUNflower analyses
of the effect of differences in traffic flow levels and of economically disadvantaged
groups on fatality rates by obtaining detailed data for cities of different size in
several countries.
6.3. Role of cultural factors
between France and Great Britain
policies addressing key accident ca
factors could be speeding, drinking and d
factors explain 80 to 90% of the gap in risk between France and Great Britain. He
further claimed that if risk factors associated with these variables in France had the
values they had in the UK, the number of fatalities in France might be reduced by
95
about half. This was based on the multiplicative effect of a reduction of one third in
speed related fatalities, by 20% in alcohol related fatalities, by 15% in seatbelt use
related fatalities and by 5% as a result of differences in urbanization. The question
mains why do these differences in behaviour occur at the national level, and are
enda, history of established safety cultures, well defined distribution of
sponsibilities between road safety 'actors', development of control and
e approach to other countries
such as France the inclusion of these issues is desirable.
oad users. This in turn will probably require
addressing issues such as the appropriate balance between safety and mobility, and
for freedom and autonomy in their
choices. For many policies which involve giving up such freedoms, public belief in
.g. Wong et al.,
006; Elvik, 2001) and they can also be expected that to play an important role at
in Great Britain.
There is sufficient evidence in success stories about the implementation of
innovative policies in urban traffic planning and organization. The wide implemen-
re
they true for all regions within a country?

Delorme (in Delorme & Lassarre, 2005) argues that it is not sufficient to address
only the accident and policy outcomes; it is more important to attempt to model
'behaviour shaping mechanisms'. By example, he quotes six areas of potential
difference between France and Great Britain: ranking of road safety policy on the
national ag
re
enforcement, development of public accountability, and existence of transparent
goals and information, parliamentary debate and active organizations. He suggests
that the stage of development of these factors was similar for the original three
SUNflower countries, and thus comparisons could be made without addressing
them too closely. But he argues that when extending th

The idea that an efficient organizational/policy structure is prerequisite for the
introduction and successful implementation of policies has gradually gained in
popularity in recent years (see e.g. Aeron-Thomas et al., 2002; Bliss & Breen, to be
published). The problem is how to consistently identify the key factors of a road
safety management system and how to measure them in a way that will show their
influence on safety outcomes.

Whatever approach is adopted, it is likely to need to address the influence of social
norms and legal and regulatory standards on the key actors involved in both the
safety policy decision-makers and the r
the public response to coercion and desire
the credibility of the policy and the motives of the policy makers is crucial. Public
willingness to spend resources on such policies, and their view of the acceptability
of their allocation will also be a major factor.
6.4. Factors affecting measures and programmes
Several authors have identified setting policy targets and strong political leadership
as determinative factors for effective road safety management (e
2
the sub-national level. Policy programmes and actions with a large road safety
improvement potential typically concern the national level. However, many of these
policies are implemented at the sub-national level by regional and local authorities.
Their commitment and professionalism are crucial for successful implementation of
particular programmes and measures. Delorme & Lassarre (2005) concluded in their
work on the comparison of France and Great Britain, that the work of local officers in
Great Britain with designated responsibilities and allocated resources is one of the
major reasons for the success in improving road safety

96
tation of 30 km/h and 60 km/h zones had a positive impact on road safety (Wegman
& Aarts, 2005). Road and traffic policies and traffic calming schemes have also a
proven to have a direct effect on road safety outcomes.
Thus we may expect that the existence of region d safety p
the attribution of necessary resources for their impleme
impact on road safety at the sub-national level.
6 oach
T f the reasons for the observed outcomes, SUNflower used
c similarities and differences in policy d the outcomes o
a with either road user group r strategic acciden
causation factors. Whilst these case studies sought to explore policies, performance
of the 'layers' of the safety pyramid. Confounding factors include the
ultiplicity of safety actions contributing to the final outcomes, and the difficulty of
uantifying the management actions and their effectiveness. Implementation of
n at the national level can only reflect the average penetration of
policies throughout the country.
Koornstra et al. (2002) included an attempt to explain the observed changes in
studies also showed the extent to which different combinations
f policy were associated with different final outcomes as for example with drinking
or in driver compliance with policies.
6.4.2. Scope for extended SUNflower analyses
Koornstra et al. (2002), in their case study on low cost infrastructure improvements,
attempted to address some of the issues raised above by tracking the historical
development of infrastructure improvement policies, and describing the tactical,
funding and operational activities associated with them in each country. But
quantifying these factors in national terms is extremely difficult. Similarly, defining at
a national level the structural factors, such as transport policy, network
characteristics, and traffic flows, against which the local programmes are being
developed is very difficult.

It is likely that at a more local level a more coherent relationship can be developed
between such structural factors, the actual level of safety policy introduced
(particularly in relation to physical engineering changes) and the way in which the
decisions were made and the influence of organization systems on these decisions
and their effective implementation. There is likely to be much more variation

al and local roa olicies and
ntation have a positive
.4.1. The SUNflower appr
o gain understanding o
ase studies of the an f
ccident groups associated s o t
indicators, and casualty outcomes, they failed to establish any clear causal links
between each
m
q
policies also differs between areas within each country, and any quantification of
implementatio

fatality numbers over the previous 20 years in the three SUN countries in terms of
the changes in exposure of the main road user groups and the trends in
performance indicators reflecting the influence of vehicle design, drinking and
driving, seatbelt wearing and road engineering programmes.

The SUNflower case
o
and driving in the balance between setting legal BAC levels, frequency of testing,
and the level of penalty for offenders. But it was not possible to assess how far the
difference in the policy combinations had contributed to the final outcome. Similarly,
driver opinions from SARTRE which differed between countries were recorded, but
these can only be expected to give a general impression of why different choices
are made either by policymakers
97
between local areas in the adoption of specific policies. Analysis at this level w
also allow the influence of specific 'safety champions' to be reflected in the loc
utcomes, either indi
ill
al
vidual or of organizations. A 'good safety outcome' might be
chieved by a variety of different sources.
se issues can be
troduction of cycling facilities and enforcement or of education initiatives may differ
substantially betwe
6.5. Regional comparisons
o date, much has remained unexplored across Europe in sub-national analysis of
ccident data. Despite the availability of local statistics in most European countries,
tics in the application of local policies by authorities. This is
r example done in France, Switzerland, Belgium, United Kingdom and Spain.
considered. This
is a purely descriptive approach, which for the moment probably lacks an analytical
crude road mortality rates, despite giving
o
a

The SUNflower pyramid already includes a layer in which the
captured and described. However, this layer needs to be substantially elaborated to
include qualitative or quantitative measures of organizational structures, social
norms, and public response to the issues discussed above. Innovative scientific
approaches would need to be taken into account in order to deal with them
consistently.

While some policies will inevitably be controlled more at the national level,
ngineering measures such as the implementation of 30km/h zones or the e
in
en at the sub-national level.
6.5.1. Current practices and applications
T
a
few analyses use the data in a systematic and consistent way. More specifically,
only a relatively small number of countries systematically monitor and reflect
regional accident statis
fo

The EC has recently incorporated a GIS application in the CARE database, which
allows mapping and monitoring of the development in road fatalities per head of
population at sub-local level. The NUTS-1 to NUTS-3 regions are
and interpretation framework. Maps of
some guidance, cannot serve alone as evidence for policy actions due to the
number of underlying structural factors which has an impact on both level and trend
of road mortality. The variation in road mortality rates across regions in 25 EU
countries in 2004 is shown in Figure 6.5 for the NUTS-2 and NUTS-3 regions. It is
based on data provided by road and policy authorities of EU countries (Eksler et al.,
2008a).

98
1 ,1 - 5 , 4
5 ,5 - 8 , 9
9 ,0 - 1 2 ,8
1 2 , 9 - 1 8 , 6
1 8 , 7 - 3 2 , 4
N U T S 2
A o r e s (P T)
C a na r ie s ( ES)
Ma d r e
Fond : R RG GIS databas e

1 , 0 6 - 3 , 3 3
3 , 3 4 - 8 , 3 9
8 , 4 0 - 1 4 , 1 0
1 4 , 1 1 - 2 4 , 9 5
2 4 , 9 6 - 3 2 ,3 9
N U T S 3
A o r es (P T)
Ca n a r ie s (E S)
M ad re
Fond : R R G GIS databas e

Figure 6.5. Road mortality rates in EU-25 countries in 2004.
99
In France, a local accident indicator (IAL) developed by Chapelon (2002) has
recently been used for the evaluation of the effectiveness of local road safety
policies. It compares the relative fatality risks of 100 French departments, defined as
the average value of fatality risk per road type. This methodology is discussed in
more detail in Appendix 9. Although this comparison is certainly more reliable than
the one based on mortality rates, its weakness is the use of a road classification.
This reflects the responsibility of authorities for its maintenance, rather than its
design and passive safety characteristics.

Switzerland monitors the road safety performance of the three linguistic regions and
the local policies also reflect the road safety performance indicators measured for
each individual region (Siegrist et al., 2006). Also in Belgium, where the two
linguistic regions have a certain authority in policies related to traffic safety, the
accident statistics are analysed at a regional level. The road mortality rates are
generally used for a comparison of the performance for regions. In the United
Kingdom, annual data is published for each of the 87 unitary authorities in England,
and for Wales, Scotland and Northern Ireland separately, enabling the calculation of

roughton & Buckle (2006) also compared the casualty rates for 88 English local
l. (1996) developed a model to investigate the high pedestrian casualty
rate in Scotland compared with the rate in English regions. The model used was

sults showed that when these variables were taken into account, the Scottish
pedestrian casualty rate was similar to that obtained by applying the model to
English regions. In Spain, the accident outcomes of Catalonia region have been
subject of interest in recent years (Hayes et al., 2006).

Harland (1999) subsequently used a similar model to explore reasons for the
relatively high observed casualty rate per head of the population in the North West
region of England in comparison with other English regions. The analysis compared
regions and also compared differences in district rates within the North West region.
Observed casualty rates were compared with predicted rates for each of five road
user groups to identify the situations in which the observed rates were higher than
those predicted by the model.

As part of their regular monitoring of national casualty trends in relation to the
national casualty target, Broughton & Buckle (2006) showed the variation between
English Regions in both fatalities per head of the population and in the development
of casualty trends between 1998 and 2005.

ity
ensity, as indicated earlier by Ekslers analyses.
fatality rates by population and by number of licensed vehicles. Fatality numbers are
also recorded by road class for eleven Government Offices in England.

B
authority areas which had been defined as deprived areas. The fatality rates per
head of population in these areas are 30% higher for adult fatalities and 60% higher
for child fatalities than in other local authority areas. Population density is likely to be
higher in the deprived areas, which may normally be expected to reduce the fatality
rate. In the case of the deprived areas, this effect may be far outweighed by factors
such as the lack of off-road playing space for children, heavier traffic volumes and
the greater difficulties in child supervision due to factors such as family structure.

Harland et a
based on population, car ownership, and the amount of road space available. The
re
Analysis of data for Dutch regions (Figure 6.6) shows a strong link between mortal
rates and population d

100
Mortality rate for different Dutch regions
Lb
Nb
Gr
RH
RR
Fl
1000 1500 2000 2500
RE
Ov
Dr
Zl
Gld
Fr
8
10
12
i
n
h
a
b
i
t
a
n
t
s
RAN
Zh

p
e
r

1
RA
Ut
Nh
RT
6
0
0
,
0
0
0

RU
4
F
a
t
a
l
i
t
i
e
s
2
0
0 500
Inhabitants per km
2

Figure 6.6. Relations between population density and road mortality in the 19
Netherlands regions (provinces and metropolitan areas) in 2004.


wever, for the four metropolitan areas, Amsterdam (RA), The Ho Hague (RH),
een mortality
argets.

SUNflower type analyses, it would be useful to
Rotterdam (RR), and Utrecht (RU), which are heavily urbanized regions with
population densities between 1,000 and 2,500 inhabitants per square km, there
seems to be relatively little effect on the mortality rate. Also, among the regions with
ensities below 1,000 inhabitants, there are large differences betw d
rates of regions with similar densities, even although the effect of population density
is a strong factor. This suggests additional factors need to be investigated in order to
more fully explain safety performance.

In Britain, it has recently been considered whether different performance targets
should be set for different regions of the national road network. Analysis of recent
progress compared with casualty rates at the start of the period being assessed,
and the traffic growth during the period suggested that inclusion of these factors
would provide a better basis for setting different targets for each region than for
simply setting the same target. For motorways (M and A(M) roads), only the starting
rate was an important factor; for other main roads both starting rate and traffic
rowth were used to identify potential separate t g

An overview of articles on regional analyses and their focus is presented in
Appendix 10.
6.5.2. Avenues for further work
There would appear to be scope for an extension of current and past work on
modelling using data from a large number of regions.

The differences indicated by population density reflect a complex set of factors that
are likely to be associated with this density e.g. numbers and lengths of trips,
choice of transport mode, and road conditions (e.g. traffic flows) faced by each
transport mode. These factors will in turn also be affected by the quality and quantity
f the networks provided for each mode. To gain deeper understanding of the issue, o
and how it might be included within
101
explore the immediate effect of changes in these factors, and to investigate their
interdependency. This could be attempted at an aggregate level - for example, by
adding more variables to the type of model already used by Eksler & Lassarre,
2008). Alternative models, such as that used in Harland (1999) focussing on
fatalities to individual road user groups, which used local car ownership data, could
also be considered.

More directly, an attempt might be made to predict travel kilometres, both vehicular
and pedestrian, that might arise as a result of population factors, income, and
transport provision, and then relate the expected number of fatalities to travel
kilometres, taking into account network quality and road user behaviour. This could
shed some light on the impact on road safety of modal split development in time.
Recent work of Stipdonk & Berends (2008) analysing road safety development in
disaggregated road user groups unveils some surprising differences in safety
development per transport mode which are not shown by studying only the overall
trend for all road users.

Extended analysis at national and at regional level would make use of data in
national databases where possible. A temporal statistical model could be run in
order to investigate the development of road safety outcomes in time. A country
trend could be considered at the same time. Combination of data on fatalities and
injuries and the use of a Full Bayes spatiotemporal model is preferred.

A disaggregated model for different transport modes and roads (speed, allowed
users), could also be considered. When thinking about the possible explanatory
factors which could be introduced to an explanatory model, one should consider the
relationships that exist between them. This is illustrated by the correlation matrix for
road link data (Figure 6.7). Thick lines indicate strong correlation ( > 0.6) and thin
or no lines indicate weaker correlation among variables. For example, the number
of lanes and the presence of a central island correlate strongly (Greibe, 2003).

Correlation between urban road design features

Figure 6.7. Correlation between various factors of urban road design features.


102
Modelling accidents for road links is less complicated than for junctions, probably
due to a more uniform accident pattern and a simpler traffic flow exposure, or due to
lack of adequate explanatory variables for junctions. Explanatory variables
describing road design and road geometry proved to be significant for road link
odels but less important in junction models. The most powerful variable for all
to traffic
volumes and various road design parameters. Accident Modification Factor models
aim to show more directly the extent to which changes in the design of individual
road features are likely to affect accident frequencies both in terms of individual
accident types and total accident frequency. Both approaches are complementary to
each other. If this approach is followed, attempts have to be made to scale up
results from individual groups of roads to road networks in regions or cities.

An alternative for regions could be to develop their safety performance profile as a
combination of the performance of the major cities within the region and an
assessment of the performance of the main intercity road network within the region.
However, the more statistical approach considered above will probably be more
appropriate for comparing regions. Comparison would most fruitfully focus on
regions where the regional authorities are the main planners and implementers of
safety policy. Dutch data might provide a useful focus for further regional analysis.
6.6. City comparisons
6.6.1. Current practices and analyses
Analysis at the regional level allows a more controlled environment to be studied
than aggregating all regions into national statistics. But most regions will still consist
of a mixture of urban and rural environments, which have clearly different road user
groups and traffic conditions as well as different population densities. One way of
reducing these confounding factors is to limit the comparison to areas having more
or less consistent characteristics, such as metropolitan regions, or to large cities.

Much work has been done on urban safety management principles, and the EC
DUMAS project included inputs from several European countries (Lines, 1999). The
main focus for urban safety management has been a combination of engineering
and speed control measures, but in the Gloucester Safer City project (Mackie &
Wells, 2003), for example, much attention was also given to the development of a
safety strategy and the organizational structure required to ensure its acceptance by
the public.

Earlier EC studies (ADONIS, WALCING and PROMISING) looked in depth at the
physical and organizational background of improvements in the safety of vulnerable
road users. ADONIS, for example, examining three major cities (Copenhagen,
Amsterdam and Barcelona) concluded that lack of safety was an inhibiting factor to
cycle use in Barcelona, but not in the other two cities.

Within each country there are also clear differences in the policies adopted by
different cities. In Britain, London has been able to introduce transport policies, such
m
models was motor vehicle traffic flow (Greibe, 2003).

Recent research literature distinguishes two approaches to modelling the safety
performance of individual roads (Eenink et al., 2008). Accident Prediction Models
attempt to define the overall risk for particular road types in relation
103
as congestion charging, which do not exist elsewhere. London
substantive development of cycling networks and 20mph (30k
also carried out a
m/h) zones. Other
British cities, e.g. York, Hull, Cambridge are also known to have pursued more
ser groups using them in different ways. This usage is reflected in different
peed of motorized vehicles to reduce the
d categories only
radical policies for residential areas, speed management, and provisions for
vulnerable road users. Comparison of policies, safety management organization and
casualty trends within these cities may yield interesting associations. Similarly Paris
and other French cities have been developing and applying specific road safety
policies aimed at protecting vulnerable road users and encouraging the use of
cological friendly modes of transport. e
6.6.1.1. Safety of different roads within cities
Different roads within cities are designed for different traffic functions, with different
oad u r
total safety outcomes and different mixes of casualty type on each road type. The
biggest difference in road function is between main traffic arteries carrying large
volumes of motorized traffic, and residential access roads where many safety
ractitioners attempt to limit the flow and s p
risk for pedestrians and cyclists, particularly children, who are likely to use these
roads. Roads distributing traffic between these two road categories often have
mixed functions, unless their layout and use are strictly controlled.

Urban safety analyses (e.g. IHT, 1990) show that typically half of all casualties in
urban areas occur on main roads, with roughly a quarter on residential roads and a
uarter on distributor roads. Analyses on the basis of these roa q
tend to be available when detailed city analyses have been made; more generally
data relate to road classifications. For example, in Gloucester (DTLR, 2001), 57% of
casualties occurred on Class A and B roads (which made up 17% of road length),
20% of casualties were on Class C roads (11% of length) and 23% of casualties on
lower class roads (72% of road length within the city).

Data from the Netherlands (Van Schagen & Janssen, 2000) illustrate the difference
between the safety characteristics of urban distributor roads (including main
arteries) and urban residential roads (Table 6.2).

Characteristic Urban distributor Urban access
Fatality risk per billion vehicle kilometres 14.5 7
Number of fatalities per year 308 60
Share of national road length (%) 13 35
Share of car kilometres (%) 18 7
Share of fatalities (%) 28 6
Share of in-patient casualties (%) 43 12
Table 6.2. Safety related characteristics of urban roads with different functions in the
Netherlands.


Table 6.2 again shows the greater role of the larger roads in the fatality toll of the
urban areas.
104
6.6.1.2. Modal influences
Pfundt & Meewes (1986) and Brhning (1986) both showed that pedestrian and
bicycle accidents made up over 40% of the total accident cost in German towns.
his proportion was relatively constant for all urban areas above 20,000 inhabitants,
d for between 35 and 85% of all road deaths in capital cities (Figure 6.8).
he availability of public transport and dedicated infrastructure together with climatic
T
although the part of the total resulting from pedestrian accidents increased with town
size, while that from cyclists increased with town size up to 50,000 and then
decreased slightly.

Data for a selection of capital cities, averaged over 2004 to 2006, indicated that
vulnerable road user deaths (pedestrian, cyclist and powered two-wheeler)
accounte
T
conditions leading to differences in modal split can partly explain the recorded
distribution of road user deaths.

Road deaths per transport mode
20%
30%
40%
60%
70%
80%
90%
100%
u
s
s
e
l
s
u
c
h
a
r
e
s
t
a
t
i
s
l
a
v
a
P
r
a
g
u
e
M
a
d
r
i
d
u
d
a
p
e
s
t
L
o
n
d
o
n
V
i
e
n
n
a
W
a
r
s
z
a
w
D
u
b
l
i
n
B
e
r
l
i
n
P
a
r
i
s
O
s
l
o
P
e
r
g
e
50%
c
e
n
t
a
0%
10%
B
r
B B
r
B
Pedestrian Cyclist PTW rider Car occupant Others

Figure 6.8. Road user deaths as a percentage of all deaths in a sample of capital
cities (ETSC, 2008).

6.6.1.3. Casualty rates by type of local authority area in England
r thousand
habitants) from 2000-2002 for these different types of area are compared in Figure
Broughton (unpublished) compared casualty rates for local authority areas in
England, categorized by a classification system produced by the Office of National
ONS, based on a Cluster Analysis of the data collected by the 2001 Census. This
identified groups of local authorities with common characteristics, as expressed by
42 Census variables. The average annual rates (casualties pe
in
6.9. The categories include both different types of urban area, and more rural areas.

105
Casualty rates for different local authority areas
All r o ad u ser s Pe destr ian s
Lon do n Ce ntr e
Lo nd on Co smo po lit an
Ce ntr e s with In dust ry
Re gi on al Ce nt re s
L ond on Su bu rb s
In dust ria l Hint er la nd s
M an ufa c tur in g Town s
Th ri vin g Lo ndo n Pe r iph e ry
Ne w a nd Gr owin g Town s
Coa sta l a n d C ou ntr ysi de
P ro spe ri ng Smal ler Town s
0 50 10 0 15 0 2 00 25 0
Pr osp er in g Sou the r n E ng la nd
0 5 0 10 0 15 0

Ki lle d KSI All ca sua lti es

Figure 6.9. Casualty rates for pedestrians and all road users for different local
authority areas, 2000-2002 (100 = national average for each rate).


Although the rates for all road users vary across the whole range of area types, the
rates for pedestrian casualties are more uniform, once London and major regional
centres are excluded.

The extent of motorized travel within a city is likely to be affected by the prominence
of the city within a region, and its proximity to good interurban road links. In the
Netherlands, for example, a high proportion of vehicular travel occurs on the
national network due to the relatively high density of the network.
6.6.1.4. Changes in safety performance
Changes in safety performance over the same period can differ substantially
between cities. Figure 6.10 shows the percentage reduction in killed and seriously
04-2006

injured casualties in English urban areas averaged over the period 20
compared with that averaged over the period 1994-1998. The individual outcomes
varied from a small increase in casualties to reductions of over 40%.

106
Variation in casualty reduction in English urban areas
0
2
4
6
> -20 -20 to 0 0 to 20 20 to 40 >40
Percentage reduction in KSI casualties
N
u
m
b
e
r

o
f


8
10
12
u
r
b
a
n

a
r
e
a
s
14
Figure 6.10. Variation in percentage reduction in KSI casualties in English urban
areas.


Figure 6.11 suggests, however, that the percentage reduction in casualty rates may
be associated with initial casualty rates in the period being considered; i.e.
reductions are likely to be higher if initial rates are higher. This is consistent with the
pattern for changes in major road casualty rate described in Section 6.5.1. However,
in both cases, the link is relatively weak, and likely to be only one of several factors
affecting the outcome.

KSI reduction at different initial KSI per population
20
40
60
80
o

2
0
0
4
-
0
6

(
%
)
-60
-40
-20
0
0 0,2 0,4 0,6 0,8 1 1,2 1,4
KSI per thousand inhabitants 1994-98
R
e
d
u
c
t
i
o
n

1
9
9
4
-
9

8

t
Figure 6.11. Percentage reduction in KSI casualty rates per thousand inhabitants
compared with the initial casualty rates.

6.6.1.5. Effect of city size
Figure 6.4 suggested that at high population densities, differences in density may no
longer have a major effect on mortality rates. Figure 6.12 similarly suggests city size
107
does not have a direct relation with th
appear to have an impact on the exposure and
e total KSI casualty rates, although it does
road safety performance of particular
road user groups.

KSI per population in English towns
0,4
0,5
0,6
0,7
0,8
0,9
l
t
i
e
s

p
e
r

1
0
0
0

p
e
o
p
l
e
0
0,1
0,2
0,3
0 50 100 150 200 250 300 350 400 450
Population (thousands)
K
S
I

c
a
s
u
a

Figure 6.12. Variation in total KSI casualties per head of population (averaged 2004-
2006) with urban population in English towns.


Consistent with the data for German towns in Section 6.6.1.2, a relation between
pedestrian and cycle casualty rates and town size can be seen both in English data
and Dutch data (Figure 6.13). However, the Dutch data suggests that the effect is
different for different casualty severities. While rates increase continuously for all
pedestrian injuries as urban area size increases (Figure 6.13a), for the more severe
injuries, rates initially rise with town size but then decrease for the larger cities
(Figure 6.13b).

Pedestrian casualties (all injuries) in Dutch towns
0,2
0,25
0,3
0,35
e
r

1
0
0
0

p
e
o
p
l
e
0
0,05
0,1
0 100 200 300 400 500 600 700 800
Population (thousands)
P
e
d
e
s
t

0,15
r
i
a
n

c
a
s
u
a
l
t
i
e
s

p
Figure 6.13a. Variation in Dutch pedestrian casualty (all injury) rates with town size.

108
Pedestrian casualties (killed+hospitalized) in Dutch towns
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
0
P
e
d
e
s
t
r
i
a
n

c
a
s
u
a
l
t
i
e
s

p
e
r

1
0
0
0

p
e
o
p
l
e
100 200 300 400 500 600 700 800
Population (thousands)

Figure 6.13b. Variation in Dutch pedestrian casualty (killed and hospitalized only)
rates with town size.

6.6.2. Avenues for further work
There would appear to be opportunities for comparison between cities both within
countries and between similar-sized cities in different countries. The former would
id the development of an effective ranking system within a country, while the latter
rammes;
nd 2) how the safety outputs have changed over the same period.
pected that the safety performance in a city will be influenced both by
a
would further extend understanding of what factors contribute to a good safety
performance.

It is proposed that the process of comparison should include:
casualty numbers by road type and road user group;
length of road and average traffic volumes by road type;
speed limits by road type and length;
transport and traffic policies and modal split;
any local information on behaviours such as seatbelt wearing, speeding, alcohol
involvement in accidents;
road safety measures implemented over at least a 10 year period by road type
(engineering and behavioural);
typical frequency and types of junctions by road type;
town population, population density, and possibly economic indicators of activity;
road safety plans, policies and prog
organization and funding of road safety programmes.

The primary task would be to select a set of cities with well-documented road safety
policies and compare 1) how these inputs have developed over the last 10-20 years,
a

can be ex It
transport and traffic policies, and by safety policies. In turn, traffic policies will affect,
and be affected by, the spatial layout of the town, the road network provided, and
the demand for travel by different modes and within different corridors within the city.
In the past, transportation models and urban safety management models have
109
analysed this situation for each major link in the road network. For the work
suggested here, it is suggested the city is defined in terms of the average conditions
in each road category. Some safety policies, such as enforcement, will be applied
across the whole city, although they might focus on specific problem areas, but
ifferent engineering and speed control measures are likely to be applied to different
ad within the city.
s AADT of different modes) and of the implementation of safety related
rogrammes (e.g. % of roads treated as 30km/h zones, % of road length with cycle
lanes) can be linked directly with the length of road of each type. Wider policies
(enforcement, education) can be applied across all roads, but may be targeted at
user groups (e.g. children) that are more likely to use a particular road types.

Data should be sought across all levels of the safety pyramid used in SUNflower.
Data on safety outcomes should be readily available, providing the boundaries of
the area being assessed can be defined. Many large cities include large peri-urban
developments; these are likely to perform differently than the more central city areas
and should either be excluded from the comparison or considered separately. As far
as road safety outcomes are concerned, both fatalities and injuries (injury accidents)
are of interest, due to the scarcity of fatal injury at the level of cities.

Data on safety performance indicators is likely to be less available unless local
surveys have been made. It may be possible to carry out small scale surveys of
specific kind of behaviour such as speeding, seatbelt wearing, compliance with
traffic signals etc. If this cannot be done, regional or national data will need to be
assumed. Differences in transport patterns and in implementation of policies and
programmes are likely to be fairly well known. But differences in safety management
structures and 'behaviour shaping mechanisms' are not likely to be known unless
explored through in-depth surveys.

The existence of safety 'champions' and an environment in which they can strongly
influence delivery of safety schemes has been shown to be important (Peden et al.
2004 and Bliss & Breen, to be published). The most noteworthy recent example is
the presidential intervention in France to put road safety high on the national
agenda. But there are many local examples where safety programmes have been
driven by the actions of individuals or groups. The DUMAS project, particularly in the
example of Gloucester Safer City, also demonstrated the role that efficient local
communication groups can play in interacting with the public and winning
acceptability for safety policies. Data that would be useful to collect include the
existence of a local safety strategy, the executive responsibilities of local managers,
the availability of funds, and the extent of local safety analysis and scheme
implementation skills. Much of this type of information may need to be obtained
through questionnaires and in-depth Interviews, and a safety management interview
process will need to be developed. Face to face interviews could be made in
d
types of ro

A starting point for defining performance is thus likely to be to assess separately the
roads serving different traffic functions residential access roads, distributor roads
and arterial roads. The balance between safety and mobility, and the demographics
of the population at risk, will be different for each of these road types, and thus
safety policies will have a different focus for each road type. The overall demand for
travel on each road type, reflected by the average person and vehicle flows, will
affect the impact on mobility of any changes to the road environment, and the
degree of acceptability of such policies. Both the quantification of the traffic problem
(a
p
110
participant countrie
cities in a wider ran
s, but there may also be scope for postal questionnaires to cover
ge of countries.

are affected by the factors above;
show how decisions are influenced by social norms/national policies;
re


rformance of
that
Britain contained a much wider number and range of city sizes than the Netherlands

h
Among the most populous cities within administrative city limits in Europe (as
r 2

entation of successful safety
programmes in national capital cities. However, these cities are likely to each have
f

cross the range of factors discussed above. It may be possible to collect
Output will aim to:
illustrate differences in safety management and safety outcomes;
show how these

show how public respond to policies implemented;

propose ways in which data on important factors might be collected mo
generally to allow comparison across a wider group of sub-national areas.
It would be of interest to compare both the factors affecting the performance of cities
of similar size in different countries, and also the factors affecting the pe
cities of differing size within the same country. The SUNflower study showed
or Sweden, and could provide a useful basis for within country comparisons. Given
the large number of cities within Britain, this comparison could encompass bot
similar-sized and different-sized cities.

opposed to urban or metropolitan areas), there are six cities with populations ove
millions and 16 cities with populations over 1 million. A sample of cities from these
groups would illustrate the factors affecting the implem
its own distinctive character, and both a 'safety ranking process' and a transfer o
experience might be less important than in smaller towns.

The conditions in smaller free-standing cities are likely to be very different, and the
study should seek to focus on comparing a sample of cities with populations from
100,000 to 250,000. The final choice of cities may be dictated by the availability of
good data a
more limited data for a larger sample of cities for some population groups.
111
7. Conclusions and recommendations
The SafetyNet project can be considered as a start-up of the European Road Safety
Observatory, paying attention to three different areas: collecting and analysing dat
at a macroscopic level (CARE, risk exposure data and safety performanc
indicators), in-depth-data (independent an

a
e
d in-depth accident investigation) and data
afetyNet project it was decided to include the
of
is
ework
evelopments
jurisdictions.

id)
he
f countries. This composite index is aimed to present the
her
n
a
n taken. As a
ark,
ned
from this study are such that SUNflowerNext's ambition benchmarking the safety
ted
fety community, and when reliable data becomes available.
Therefore, it is recommended to carry out this task in Europe in the near future, to
and
States.
performances

The benchmark results enable
countries or jurisdictions to learn from others as a basis for developing measures
In the SUNflowerNext project we concluded that it is better not to make comparisons
and
application (EU safety information system and data analysis and synthesis).

In the course of the present S
SUNflower approach in SafetyNet in an attempt to integrate different components
the two projects. This report reflects the work that has been done to accomplish th
task.

The aim of the SUNflowerNext project is to develop a knowledge-based fram
for comprehensive benchmarking of road safety performances and d
for a country or sub-national

The SUNflowerNext study attempts to link the different components of the road
safety target hierarchy (information from the five layers of the SUNflower pyram
and to use this to create a composite index for road safety for benchmarking t
road safety performance o
safety performance of countries in a valid and meaningful manner and, in addition,
can be used to identify how countries can learn from each other.

This study has made use of existing data that was relatively easily available. This
ensured that the study could be carried out in a relatively short time. On the ot
hand, one important concession needed to be made. Because this study is a
innovative approach using existing data only, and it turned out that relevant dat
was not always available, it was decided to do the research in such a way that all
the steps required for benchmarking a country's performance have bee
result it was decided to refrain from presenting the actual results of the benchm
as they are considered not always to be of sufficient quality. The experiences gai
performance of countries is realistic after detailing indicators, have them accep
by the road sa
widely spread the results of this benchmarking, for example on an annual basis,
to consequently make use of them for policy making in the European Member
7.1. Benchmarking of road safety
Benchmarking is a process in which countries or sub-national jurisdictions (states,
provinces, 'lnder', etc.) evaluate various aspects of their performance in relation to
other, and to so-called 'best in class' practices.
and programmes with the aim of improving their own performance.

between all European countries, but to attempt grouping comparable countries
to then compare the countries within a specific group.
112
Different approaches have been used in this study, in Chapters 3 and 4
approaches are presented. We recommend identifying the pro
these
s and cons of these
approaches taking data availability explicitly into account, and based on this, making
gful groups could be
made: safety experts were asked to group countries, countries were grouped based

of the annual fatality rates in the years 1980-2003 of
countries), and, thirdly, countries were grouped using general statistic data about a

eservation, for reasons which are given
y
a final decision on how to best carry out this grouping of countries.

Three procedures have been used to find out whether meanin
on road safety outcome indicators (grouping obtained with a Singular Value
Decomposition (SVD)
country in the most recent years based on a Multiple Correspondence Analysis
(MCA).

The results of the three methods have many points of agreement. The results are
presented in the table below with some r
before. Further effort is recommended to improve the quality of this grouping and b
adding more countries to this grouping.

Group 1 Group 2 Group 3 Group 4
Denmark
Finland
Czech Republic
Hungary
Great Britain
Iceland
The Netherlands
Norway
Sweden
Austria
Belgium
France
Germany
Irela
Poland
Slovakia
Slovenia
Greece
Portugal
Spain
nd
Italy
Luxemburg
Switzerland

Table 7.1. Indicative grouping of countries based on three different procedures to be
ors
fety performance of a country including all information as
of
ce
d
e three top layers of the SUNflower-
diate


ore
f road
cess indicator can also be used. Basically, this
used in further benchmarking of road safety performances.
7.2. Indicators for road safety
SUNflowerNext decided to develop an integral and comprehensive set of indicat
to measure the road sa
proposed in the SUNflower pyramid. SUNflowerNext distinguishes three types
indicators: the road safety performance indicator, the implementation performan
indicator, and the policy performance indicator.

The first indicator captures the quality of road safety in a country. It has been named
Road safety performance indicator. Other names such as outcome indicators an
product indicator are also used. In SUNflower th
pyramid are included: final outcomes (numbers of killed and injured), interme
outcomes (such as the safety performance indicator), and social costs.

For a meaningful comparison of countries, numbers of people killed or injured are
typically 'normalized', resulting in fatality rates, e.g. fatalities per inhabitant, vehicle
type, or kilometre travelled. Besides, the comparison may specifically concern m
vulnerable groups of road users, e.g. pedestrians, cyclists, motorized two-wheelers.

The second type of indicator specifies the quality of the implementation o
safety policies: the Implementation performance indicator. For this implemen-
tation quality indicator the term pro
113
indicator follows a vertical line in the pyramid linking 'safety measures an
programmes', safety performance indicators and numbers of kille
d
d and injured
people.
of
tween the (three) different layers of the safety pyramid:
between the policy changes ('safety measures and programmes') and the changes
ugh
ade in the development of safety performance indicators
within the SafetyNet project, it was necessary to conclude that the possibilities for a
till
ns,
evelopment and data collection.
ty:
policy
cator. Here SUNflowerNext distinguishes two components: the quality of

strategy, more specifically
lly
al
sed
and diagnosis of road safety problems, vertical
al
on
policies. However, it must be concluded that as yet only little
d to

fety Observatory. A good understanding of why European

ds a composite road safety performance index
The pros and cons of working with composite indices are rather well known and are
ristics:
st
o
evelopment Index, the Environmental Sustainability
Index and the Overall Health System Index. Based on these examples it was
The purpose of SUNflowerNext is to explore a composite road safety performance
index and to research the similarities in the behaviour of basic indicators and

Implementation performance, in general, can deal with different components
causal relationships be
in safety performance indicators (SPIs), and between the changes in safety
performance indicators and changes in the number of casualties. However, altho
much progress was m
systematic and comprehensive measurement of all these relationships are s
limited. Further research is needed on this indicator, both in terms of definitio
concept d
The third type of indicator deals with the quality of policies to improve road safe
the Policy performance indicator. This indicator sometimes is called a
output indi
conditions (strategies, programmes, resources, coordination, institutional settings,
etc.) and the quality of action plans and individual (counter)measures) in the
perspective of the ambitions expressed in road safety targets
Policy performance is about the quality of road safety
about the quality of road safety plans and the conditions for successfu
implementing road safety measures and programmes. Examples are institution
arrangements, budget, quality of professionals, application of evidence-ba
knowledge, sound analysis
cooperation between different tiers of government, etc. Different internation
studies summarize the demands for the effective development and implementati
of national road safety
information from European countries is available, and it is strongly recommende
include this type of information in international data collection systems, such as the
European Road Sa
countries are making road safety progress is impossible without information on
Policy performance.
7.3. Towar
There are several reasons why it is attractive to combine all information in one
indicator, a so-called composite index. A composite index includes all components
of the SUNflower pyramid, more specifically the three types of indicators.

discussed in this report. Three words can summarize the main characte
'simplification, quantification and communication'. Road safety will not be the fir
policy field to successfully attempt capturing performance in one single value. T
mention a few: the Human D
decided to also explore the opportunities for a composite index for road safety.

114
countries. The composite index will enable a ranking of countries according to their
safety performance.

The choice of

indicators and their definitions has a preliminary character. As was
mentioned earlier, the choice of indicators used was influenced by the direct
that
tudy
Weights based on statistical models were used to combine the basic indicators into
actor
at
ollinear to form a composite index that captures as much of common
information among sub-indicators as possible. The analysis was made on the data
re
ombined safety indicator, and clusters/
groups of countries with similar values of the combined indicator. The composite
ty
nd, concurrently, to explore the similarities in the behaviour of basic
by
es,
to their
t
r
rmance


comes, both part of the road safety performance indicator, are not

ation from the different components of the road safety

mising and seems to be preferable to
e
d
d
U
y
availability of data. In addition, the assessments that were carried out suggest
the final results that are presented are not the main value of its exercise. This s
makes clear along which lines a composite index can be designed.

a composite one. Both Principal Component Analysis (PCA) and Common F
Analysis (FA) weighting were examined. Both methods group together indices th
are c
collected for 27 European countries. Five trials of creating a composite index we
performed, where each trial produced a c
indicator enables us to rank the countries in accordance with their safe
performance.

The purpose of our analysis was to create a composite road safety performance
index a
indicators and countries. It was demonstrated that both tasks cab be realized
means of the statistical weighting methods applied. The composite indic
estimated by several methods, enabled us to rank the countries according
safety performance.

The analysis revealed that the countries' ranking based on a composite index is no
necessarily similar to the traditional ranking of countries based on mortality rates o
fatality rates/risks only. We believe that adding information on policy perfo
and implementation performance to the ranking process improves the results
beyond the traditional methods and makes them more comprehensible.
Furthermore, it was observed that the indicators belonging to the final outcomes and
intermediate out
uniform in their behaviour. Indicators which were found to be more consistent and
termed 'core set of basic indicators' are recommended for future uses.

The general conclusion is that the design of a composite road safety performance
index in which relevant inform
pyramid has been captured and weighted is realistic and meaningful. In addition,
such an index gives a more enriched picture of road safety than a ranking only
based on data on mortality or fatality rates, which is normal practice at present.
Grouping countries in this process is pro
simply ranking countries. Before defining the SUNflower road safety performanc
index and actually applying the results to policy making, two improvements shoul
be made: develop indicators for the Implementation performance indicator an
develop procedures to make available high quality and comparable data for E
Member States. Finally, it is recommended to develop a standardized terminolog
for road safety indicators and a composite index.
115
7.4. Time series analyses
Analyses of safety developments are interesting because they may give us a better
insight in the underlying forces for these developments and, hopefully, also in the
is
to compare
n fatality rates (fatalities per 10,000 motorized vehicles) and mortality
000 inhabita
e
gests that all countries
g

, we may conclude that, although all European countries tend
are
rds, the
ety in different countries ultimately seem to
oad
d
s
ended to
oping menu-based
ods and by organizing training courses.
d the
rison of programmes and performance at the sub-national
ority
here
activity result in changes in mobility
n of the SUNflower method at sub-
s are
mes and
improved
the application and results of road
ethink
rt of the
nce
ose methods allowing reliable comparison.
effectiveness of road safety interventions. Different approaches were used in th
part of the study, among which state space modelling. The first attempt
developments i
rates (fatalities per 100, nts) was made at a macroscopic level. Although
European countries do have a remarkably different history when it comes to th
development of fatality rate vs. mortality rate, our data sug
seem to be moving to the same position. However, leading countries in the field of
road safety generally keep ahead of the other countries, albeit with decreasin
advantage.

Three types of disaggregate developments are compared (age, traffic mode and
road type). For this comparison countries were firstly grouped. Inspecting the results
of the analyses
towards the same aggregated or macroscopic level of road safety, there
important differences between the individual countries as well as between groups of
similar countries in terms of how they reach this level of road safety when
considering their focus on avoiding special types of accidents. In other wo
general policies of improving road saf
move towards the same safety level, but for different countries that level of r
safety is achieved at a different pace and in different ways.

Time series analysis methods have been available for many years. More advance
methods are being developed as of recently. These methods offer new opportunitie
and more reliable results; however they require a good insight. It is recomm
introduce these methods in road safety research by devel
applications for these meth
7.5. SUNflower at regional or city level?
SUNflower analyses to date have focussed on national comparisons an
lessons that could be learned by individual countries from comparing their road
safety policies and resulting performance with that of other countries. A similar
scope exists for compa
level. In principle, such comparisons can be made at any level where an auth
has the possibility to pursue its own safety programme and policies or w
developments in motorization or in economic
levels of different transport modes, resulting in changes in safety outcomes.

There is potential for innovation in terms of knowledge and road safety improve-
ments which can be reached by the applicatio
national level, which has so far been left unexplored. The potential beneficiarie
not just the local administrations who prepare and apply the program
policies at the local level, but national administrations also benefit by
understanding of structural factors influencing
safety policies. The application at sub-national level may however require a r
and an adjustment of the method applied so far at the national level. This pa
study was intended to explore the potential for comparisons of safety performa
at a sub-national level and to prop
116

There are two basic reasons for comparing safety performance of sub
jurisdictions. In the first place, a ranking of relative performance of each are
-national
a will be
very useful for comparison within countries. To ensure a meaningful process, factors
at could enable safety practitioners to achieve similar safety improvements need
to be included within the ranking process. This might be done by developing a
model of safety outcomes that includes these factors. The second reason is to
provide better understanding of the factors affecting safety improvement, so that
safety practitioners can achieve more effective programmes. This requires greater
focus on understanding how the effects of programmes are modified by the nature
of the safety problems faced by each area. Lessons can not only be learned from
comparison between areas within countries, but also from comparison between
similar areas in different countries. This can probably be more effectively achieved
by case studies comparing small groups of areas, than by attempts to model a wide
range of areas.

There are two ways to make any statistical sub-national comparison meaningful and
reliable. One is to compare areas that are similar in terms of their physical structure
such as road network configurations, climate, relief, mobility patterns, modal split,
and attitudes towards road safety. The second is to try to adjust for as many of
these factors as possible. The first way limits the number of areas that can be used
for comparison, the second is sensitive to data reliability and availability, and implies
the use of sophisticated statistical methods. This could make the comparison less
understandable for a broader public and might bring about the risk of the use of
inappropriate and confounding factors.

An alternative approach, consistent with the SUNflower case studies, is to explore
the factors that are present or absent in situations with a good or a poor safety
performance. This approach allows factors that may be difficult to define in
numerical terms, such as organizational efficiency or the presence of constraints on
the implementation of safety policy, to be considered alongside the numerical
characteristics of an area. The outcome of such an approach will be less quanti-
tative but allows a potentially wider range of factors to be considered.

Analysis at a regional level allows a more controlled environment to be studied than
aggregating all regions into national statistics. However, most regions will still
contain a mixture of urban and rural environments, which have clearly different road
user groups and traffic conditions as well as different population densities. One way
of reducing these confounding factors is to limit the comparison to areas that have
more or less consistent characteristics, such as metropolitan regions or to large
cities.

The results of this sub-national comparisons are considered that interesting that it is
recommended to continue this work in an international/European project. In addition,
it is also recommended to use a different approach for studies at both the regional
and the urban level.

For regions

The SUNflowerNext study of applying the SUNflower methodology on regions
clearly illustrates the added value of this new work. Different research questions
have been addressed and important factors influencing the safety performance of
regions have been identified. More attention seems to be required for factors and
th
117
developments that are captured in the bottom layer of the pyramid: 'structure and
culture'. We recommend a statistical approach. Practically all structural factors are
mutually correlated and if they are used simultaneously this should be done with
care: otherwise analyses could result in confusion rather than in understanding.

For cities

We recommend further studies for comparison between cities both within countries
and between similar-sized cities in different countries. The first type of comparison
will aid the development of an effective ranking system within a country, while the
latter will further increase our understanding of what factors produce a good safety
performance. The primary task of this study would be to select a set of cities with
well-documented road safety policies and compare how these inputs have
developed over the last 10-20 years, and how the safety outputs have changed over
the same period. For this comparison we recommend to use a case-study approach.


118
References
AAA (2007). Improving traffic safety culture in the United States. The journey
velopment index (RSDI). Development
f an international index to measure road safety performance. Linkping studies in
S c echnolo entiat hesis, . 1174 epartm t of S ce an
T hno inkping vers
Bijleveld, F.D., Comma ur, . m . M
b ed ement o nt li In al
Royal Statistical Socie vo r 5
Bliss, A. & Breen, J.M. (to be publish
m nag A road s r decision- a ti W
B k, ton D.C pre
Broughton, J. & Buckle,
r ucti . TRL R rt 6 n se a T w
B sh
Broughton, J. & Buckle,
r cti t - 2004 . T o r R or
C th erkshire
Brhning, E. (1986). Unflle innerhalb von Ortschaften - Karakteristische
u rsc in abhan eit g : e r
A C 'The l of r t J 6, V .
Champion, A.G. (2001). m e an
u n s: conseq es f iz po n u ity
p la dies, 67
Chapelon, J. (2002). Un nouvel indicateur d . In: C
s sti r. 103, S mb
Christie, N. (1995). Th h r p n e
environm factors in their accid ts. Report PR
Laboratory TRL, Crowt ne, e
C m J.J.F. & pm n c ta e
series analysis
forward. American Automobile Association AAA Foundation for Traffic Safety,
Washington D.C.
Adriaanse, A. (1993). Environmental policy performance indicators. A study on the
development of indicators for environmental policy in the Netherlands. Ministry of
Housing, Physical Planning and the Environment, Sdu Publishers, The Hague.
Aeron-Thomas, A., Downing, A.J., Jacobs, G.D., Fletcher, J.P., Selby, T. & Silcock,
D.T. (2002). Review of road safety management practice; Final report.
PR/INT/216/2002. Transport Research Laboratory TRL, Crowthorne, Berkshire.
l Haji, G. (2005). Towards a road safety de A
o
cien e and T gy, Lic e T No . D en cien d
ec logy, L Uni ity.
nde J.J.F., Gould, P.G & Koop an, S.J (2008). odel
as measur f late risk in time series with app cations. : Journ of the
ty A, l. 171, n . 1, p. 26 -277.
ed). Institutional arrangements for road safety
a ement: afety manual fo makers nd prac tioners. orld
an Washing . [In ss]
G. (2005). Monitoring progress towards the 2010 casualty
ed on target epo 43. Tra sport Re arch L boratory RL, Cro thorne,
erk ire.
G. (2006). Monitoring progress toward the 2010 casualty
edu on targe data RL Rep rt 653. T ansport esearch Laborat y TRL,
row orne, B .

nte hiede gigk der Orts rosse. In Proceedings of th 5th Inte national
TE Congress ack oad safe y', 9-13 une 198 Paris, olume 3
A changing de ographic regim d evolving polycentric
rba region uenc or the s e, com sition a d distrib tion of c
opu tions. In: Urban Stu vol. 38 , p. 657- 7.
accidentologie locale ourier des
tati ques, n epte er 2002 .
e hig isk child
en
edestria
Project
; Socio- conomic
117. Transpo
and
rt Research ental
hor Berkshir .
om andeur, Koo an, S.J. (2007). A introdu tion to s te spac time
. Oxford University Press, Oxford.
119
Delorme, R. & Lassarre, S. (eds.) (2005). LInscurit routire en France dans le
m de la compariso ternat le. Report INRETS 261. I tut Na l de
Recherche r Scurit INRETS, Arce
D riks, M 0 e g tra u T
s cial r n no o an lo E
International Transport F rum ITF, Paris.
D R (2 lo s D n ns c rn
a the D on
Durbin, J. & Koopman, S.J. (2001). T s s b s th
O ord y x
EC (2001). e to decide.
E pea i ff ff li f t p
C mu ux g
EC (2003). Action Progra Halving the num r of road
a en in ro io 1 red ns
C mis h a u E tor ne n
a Tra r
EC (2007). igures. Part
T spo te fe p m Di te l f
E gy ns o n ro
Eenink, R., Reurings, M., Elvik, R., Cardoso, J.,
A den io s d safety impact assess ec da
f sin o e P E r ro
Commis us
E er, ). o E o tive d ph
s tur p y I h, n 0-
Eksler, V. (t th . PhD
t is, i
ksler, V. & Lassarre, S. (2008). Evolution of road risk disparities at small-scale
m. In: Journal of Safety Research, vol. 39, nr. 4, p. 417-427.
Eksler, V., Lassarre, S. & Thomas, I. (2008a). The regional analysis of road mortality
in Europe: A Bayesian ecological regression model. In: Public Health, vol. 122, nr. 9,
p. 826-837. doi:10.1016/j.puhe.2007.10.003
Eksler, V., Lassarre, S. & Thomas, I. (2008b). Belgian regions: Federated or divided
in road fatality risk evolution? In: 48th Congress of the European Regional Science
Association ERSA, 27-31 August 2008, Liverpool.
Elvik, R. (2000). How much do road accidents cost the national economy? In:
Accident Analysis and Prevention, vol. 32, nr. 6, p. 849-851.
Elvik, R. (2001). Quantified road safety targets: an evaluation methodology. TI
Report 539/2001. Institute of Transport Economics TI, Oslo.
iroir n in iona nsti tiona
sur les Transports et leu uil.
er H.M. & ak, P. (2 07). Und rreportin of road ffic cas alties. IR AD
pe report. O ganisatio for Eco mic Co- peration d Deve pment O CD /
o
TL 001). G ucester afer city. epartme t for Tra port, Lo al Gove ment
nd Regions TLR, L don.
ime serie analysi y state pace me ods.
xf Universit Press, O ford.
White paper
n Comm
: European transport policy for 2010;
ssion, O
Tim
he Euro uro ice for O icial Pub cations o ean
om nities, L embour .
European Road Safety
t victims
mme:
0: A sha
be
ibility. ccid the Eu pean Un n by 20 respo
om sion of t e Europe n Comm nities C C, Direc ate-Ge ral for E ergy
nd nsport, B ussels.
Energy and transport in f 3: Transport, Chapter 3.6 Means of
ran rt; Chap r 3.7 Sa ty. Euro ean Com ission, rectora -Genera or
ner and Tra port in c -operatio with Eu stat.
Wichert, S. & Stefan, C. (2008).
cci t predict n model and roa ment r ommen tions
or u g these t ols. Deliv rable D2 of the RI CoRD-IS REST p oject. Eu pean
sion, Br sels.
ksl V. (2007 Road m rtality in urope: H w sensi is it to emogra ic
truc e and po ulation d namics? n: IATSS Researc vol. 31, r.1, p. 8 88.
o be published). Road mortality in e EU: a regional approach
hes n press.
E
level: example of Belgiu
120
Elvik (2008). Dimensions of road safety problems and their measurement. In:
Accident An and tion, 0, nr 1200
E ik, R. & , K . B th f ef ass t tool
r d safe W age E re roj EBU
T I rappo 005 te o or ic
E SO (20 ua ica 20 pe Sa serv
ERSO. www.erso.eu
ETSC (2001 ort Safety
C ncil E uss
E C (20 eth cal h t al r ty European
Transport Safety Council ETSC, Bru
E C (20 du ths ink PI 5 B nd d
E pean rt ou C, s.
E C (20 re at PI 4 B nd urope
T sport ou C, ls.
E C (20 ing ita PI 11 an rt Sa
C ncil E uss
G ham, J tep A. De ng ct ati
child pede su En : A An d
n , p. 1 4.
Greibe, P. (2003). Ac red od an n: Ana
and Prevent
H ert, A lma Vis ds ). R ety anc
indicators: . D le D e pro ety opea
Commissi se
Harland, G. (1999). C rat N st. ort nspo
R arch ory row Be
arland, G., Bryan-Brown, K. & Christie. N. (1996). The pedestrian casualty problem
in Scotland: why so many? The Scottish Office Central Research Unit Research
Report. Stationery Office, Edinburgh.
Harvey, A.C. (1989). Forecasting, structural time series, and the Kalman filter.
Cambridge University Press, Cambridge.
Hayes, S., Serrano, S., Pags, L., Zori, P., Handanos, Y., Katscochis, D., Lemonde
de Macedo. A., Cordoso, J. & Vieira Gomes, S. (2005). SUNflower+6; A
comparative study of the development of road safety in Greece, Portugal, Spain,
and Catalonia. Design & Systems Development DSD, Barcelona.
Hermans, E., Van der Bossche, F. & Wets, G. (2008). Combining road safety
information in a performance index. In: Accident Analysis and Prevention, vol. 40, p.
1337-1344.
alysis Preven vol. 4 . 3, p. -1210.
lv Veisten . (2005) arriers to e use o ficiency essmen s in
oa ty policy. orkpack 2 of the uropean search p ect ROS D.
rt 785/2 , Institu f Transp t Econom s, Oslo.
R 08). Ann l Statist l Report 07. Euro an Road fety Ob atory
.
). Transport
TSC, Br
safety perfor
els.
mance indicators. European Transp
ou
TS 06). A m odologi approac o nation oad safe policies.
ssels.
TS 07a). Re cing dea from dr driving. N Flash ackgrou ata.
uro Transpo Safety C ncil ETS Brussel
TS 07b). Inc asing se belt use. N Flash ackgrou data. E an
ran Safety C ncil ETS Brusse
TS
ou
08). Mak
TSC, Br
EU cap
els.
ls safer. N Flash . Europe Transpo fety
ra .D. & S hens, D. (2008). composi the impa of depriv on on
strian ca
351-136
alties in gland. In ccident alysis an Prevention, vol. 40,
r. 4
cident p iction m el for urb roads. I Accident lysis
ion, vol. 35, nr. 2, p. 273-285.
akk .S, Gite n, V. & , M.A. (e .) (2007 oad saf perform e
Theory eliverab 3.6 of th EU FP6 ject Saf Net. Eur n
on, Brus ls.
asualty es in the orth We TRL rep 393. Tra rt
ese Laborat TRL, C thorne, rkshire.
H
121
Hooghe (1995). Sub-national mobilisation in the European Union. In: Hayward, J.
(ed.), The crisis of representation in Europe. Frank Cass, London.
IHT (1990). Urban safety management; Guidelines from the Institute of Highways
and Transportation IHT. Department of Transport, Traffic Policy Division, Traffic
Advisory Unit, London.
IIHS (2006). Bad statistics lead to misinformation. In: Insurance Institute for Highway
Safety IIHS Status Report, vol. 41, nr. 4. Insurance Institute for Highway Safety
IIHS, Arlington, USA.
ISO (2006). Environmental management - Life cycle assessment - Principles and
rming variables for normality and linearity When,
ow, why and why not's. In: SAS Conference Proceedings NESUG 2005, 11-14
inear algebra and its applications. Addison-Wesley Publishing
Le Gals, P. (1998). Regulations and governance in European cities. Blackwell
Lin 9). The DUMAS project - Developing Urban Management and Safety.
S
u 2001. Working
nt of road
ackie, A. & Wells, P. (2003). Gloucester safer city; Final report. TRL Report 589.
Transport Research Laboratory TRL, Crowthorne, Berkshire.
framework. ISO 14040:2006. International Organization for Standardization ISO,
Geneva.
KfV (2007). Best practices in road safety. Handbook for measures at the country
level. Final report of the SUPREME project, Part C. European Commission,
Brussels.
Koornstra, M., Lynam, D., Nilsson, G., Noordzij, P., Pettersson, H-E., Wegman, F. &
Wouters, P. (2002). SUNflower; A comparative study of the development of road
safety in Sweden, the United Kingdom, and the Netherlands. SWOV Institute for
Road Safety Research, Leidschendam, the Netherlands.
LaLonde, S.M. (2005). Transfo
h
September, Portland, Maine.
Lay, D.C. (1994). L
Company, New York.
Publishers, Oxford.
es, C. (199
In: Traffic management, safety and intelligent transport systems, proceedings of
eminar D, AET European Transport Conference, 27-29 September 1999,
Cambridge,.
LTSA (2000). Road safety strategy 2010; A consultation document. National Road
Safety Committee, Land Transport Safety Authority LTSA, Wellington, New Zealand.
yten, S. & Van Hecke, E. (2007). De Belgische stadsgewesten L
Paper, nr. 14, Statistics Belgium.
Lynam, D., Nilsson, G., Morsink, P., Sexton, B., Twisk, D., Goldenbeld, C. &
Wegman, F. (2005). SUNflower+6; An extended study of the developme
safety in Sweden, the United Kingdom, and the Netherlands. Transport Research
Laboratory TRL, Crowthorne.
M
122
Morsink, P., Oppe, S., Reurings, M. & Wegman, F. (2005). SUNflower+6;
Deve application of a footprint methodology for the SUNflower+6
countri Institute Saf earch, Leidschendam, the
Netherlands.
Nardo, M., Saisana, M., Saltelli, A., Tarantola, S., Hoffman, A. & Giovannini,
(2005). Handbook on constructing com dica tho d us
guide. OECD Statistics Working Papers, 2005/3, Or on f mic
operation and Development OECD, Paris.
Noland, R.B. & Quddus, M.A. (2004). A spatially disaggregate analysis of roa
casualties in England. In: Accident Analysis and Pr , vo 6, p
984.
OECD (2002). Safety on roads. What's the vision? tion nom
operation and Development OECD, Paris.
OECD/ECMT (2008). Country reports o afe an D/E
working group on achieving ambitious road safety t oint CM
Transport Research Centre (JTRC), Paris.
http://www.internationaltransportforum.org/jtrc/safety/targets/Performance/pe
ce.html
OECD/ITF (2008). argets and the Safe
System Approach. Joint Transport Research Centre of The OECD and the
International Transport Fo is.
O D/JRC (2008 ns com dicators. Method
and user guide. Organisation for Economic Co-operation and De ent O
and the Joint Research Centre of the European Commission. OECD, Paris.
Peden, M., Scurfield, R., S Moh yd Jara & M

fundt, K. & Meewes, V. (1986). Gestaltung von Straennetzen und Straenrumen
nter dem Aspekt der Verkehrssicherheit in Stdten und Gemeinden. In:
Proceedings of the 5th International ATEC Congress 'The lack of road safety', 9-13
June 1986, Paris. Volume 3.
Saisana, M. & Tarantola, S. (2002). State-of-the-art report on current methodologies
and practices for composite indicator development. EUR 20408 EN. Joint Research
Centre of the European Commission, Italy.
Sassi, F. (2006). Calculating QALYs, comparing QALY and DALY calculations. In:
Health Policy and Planning, vol. 21, nr. 5, p. 402-408. doi:10.1093/heapol/czl018.
Schagen, I. van & Janssen, T. (2000). Managing road transport risks: Sustainable
Safety in the Netherlands. In: IATSS Research, vol. 24, nr. 2, p. 18-27.
Schmidt-Seiwert, V. (1997). Landkarten zum Vergleich westlicher Regionen. In:
Hradil, S. & Immerfall, S. (eds.). Die westeuropischen Regionen im Vergleich.
Opladen, p. 603-628.
lopment and
es. SWOV for Road ety Res
E.
posite in tors: me dology an er
ganisati or Econo Co-
d
evention l. 36, nr. . 973
Organisa for Eco ic Co-
n road s ty perform ce. OEC CMT
argets. J OECD-E T
rforman
Towards Zero Ambitious Road Safety T
rum, Par
EC ). Handbook on co tructing posite in ology
velopm ECD
leet, D., an, D., H er, A.A., wan, E. athers,
C. (2004). World report on road traffic injury prevention. World Health Organization
WHO, Geneva.
P
u
123
Siegri llenbach a . nn c , 6).
SINUS-Rep Strassenverkehr
2005. izerische at e ll g
S do & Berend . (2 i i m a
s ty . In: Accident An
p 38
S O ). Road c h c s . n
for Road Safety Research, Leidsch
S O ). Cost-benefit analysis o O s
J e OV Ins e S s d
Netherlands.
T c .G. & Fi , L. ) e s n
Bacon, Boston.
T i si, G., Gi ini, c o a
Marturano, P. (2006). e "U s t o h
b ets aly 0 S 9
S r nita, R a. [ ]
Thaler, R.H. & Sunstein, C.R. (2008). m a
wealth, and happiness le ity
Trinca, G.W., Johnstone bell, B.J., H
G.M., , A.J. & uc d ff A
c le yal Australasian College o b
UNECE dat
h // e.org/px /D v V R
h + 31+Dec be pe+of+vehi u
a Y ath=../DA A 0 e 1
V M ent, A.L. (e 0 r e r
C n sons. ive p E
Rees, W. (1996). Our ecological footprint. Reducing human
pact on the earth. New Society Publishers.
egman, F. (2004). Implementing, monitoring, evaluating and updating a road
safety programme. D-2003-12. SWOV Institute for Road Safety Research,
Leidschendam, the Netherlands.
Wegman, F., Eksler, V., Hayes, S., Lynam, D., Morsink, P. & Oppe, S. (2005)
SUNflower+6; A comparative study of the development of road safety in the
SUNflower+6 countries: Final report. SWOV Institute for Road Safety Research,
Leidschendam, the Netherlands.
st, S., A , R., C vegn, M , Niema , S. & A kermann Y. (200
ort 2006; Sicherheitsniveau und Unfallgescheh
Schwe
en im
BfU, Be Ber ungsstell fr Unfa verhtun rn.
tip nk, H. s, E 008). D stinguish ng traffic odes in nalysing road
afe
. 1
development
3-1393.
alysis and Prevention, vol. 40, nr. 4,
W V (2007 ras osts. SWOV Fact heet, March 2007 SWOV I stitute
endam, the Netherlands.
W V (2008 f road safety measures. SW V Fact heet,
un 2008. SW titut for Road afety Re earch, Leidschen am, the
aba hnick, B dell S. (2006 . Using multivariat statistic . Pearso /Allyn &
agg , F., Do ust M., Cren a, A., Cedri, C., F ndi, G., I scone, P. &
Th lisse" Sy tem for he monit ring of the use of t e safety
elts and the helm in It (2000-2 05). Rapporti ISTI AN 06/3 . Istituto
upe iore di Sa om In Italian .
Nudge. I proving decisions about he lth,
. Ya Univers Press.
, I.R., Camp aight, F.A., Knight, P.R., Mackay,
McLean Petr elli, E. (1988). Re ucing tra ic injury; global
hal nge. Ro of Surge ns, Mel ourne.
abase (2008).
ttp: w3.unec web ialog/var al.asp?ma=01_TRRoadtyp eh_r&ti= oad+ve
icle fleet+at+ em r+by+Ty cle%2C+Age+gro p%2C+Country+
nd+ ear&p TAB SE/Stat/4 -TRTRANS/02-TRRoadFle t/&lang=
is, .A. & G van ds.) (20 7). Road safety pe formanc indicato s:
ou try compari Del rable D3.7a of the EU FP6 roject SafetyNet. uropean
Commission, Brussels.
Wackernagel, M. &
im
Wegman, F. & Aarts. L. (ed.) (2005). Advancing Sustainable Safety; National Road
Safety Outlook for 2020. SWOV Institute for Road Safety Research, Leidschendam,
the Netherlands.
W
124
Wong, S.C., Sze, N.N., Yip, H.F., Loo, B.P.Y., Hung, W.T. & Lo, H.K. (2006).
Association between setting quantified road safety targets and road fatality
reduction. In: Accident Analysis and Prevention, vol. 38, nr. 5, p. 997-1005.

125

Appendix 1. Final data set with initial and imputed
values of basic indicators
For data imputations, the SAS MI pr was applied using the MCMC
method for imputation. At present, s can be used for the
MCMC imputation. This imputat sume ariate normality of variables. We
used SAS macro %bctrans to Box- nsformation in order to achieve
marginal normality of each of th ontinuous variables. (Details about %bctrans can
e found in LaLonde, 2005.) Due to the relatively small number of observations with
respect to the number of var sic or putation was carried out
separately for groups of variables. All the data transformations and imputations were
done for the data set with Malta. Th
observation was imputed using a monotone imputation. The final dataset of basic
indicators is as follows in the tables belo

9.2
only continu
ocedure
ous variable
ion as s multiv
find a Cox tra
e c
b
iables (ba indicat s), the im
out en, the missing value among Malta's single
w.
Obs Country Cod B1 B2 B3 e B4 B5 B6
1 Austria AT 88.138 174.6 99.697 54.63 05 56 0.15068 0.06575
2 Belgium BE 101.347 216.0 96.160 38.46 72 02 0.11413 0.08606
3 Cyprus CY 111.320 236.2 167.478 31.08 38 14 0.22093 0.02326
4 Czech Republic CZ 103.514 263.532 145.756 20.8043 0.19003 0.10348
5 Denmark DK 8 15 55.802 17.65 56.27 3.588 69 0.19608 0.10131
6 Estonia EE 151.837 3 175.258 12.67 89.390 16 0.27219 0.04142
7 Finland FI 63.802 136.146 53.034 20.0595 0.11873 0.11346
8 France FR 76.753 1 63.912 17.05 53.140 44 0.11361 0.03844
9 Germany DE 61.802 1 57.312 64.42 09.886 43 0.13966 0.09546
10 Great Britain UK 54.385 1 47.645 58.89 15.697 02 0.20953 0.04556
11 Greece EL 148.630 374.626 174.421 9.7707 0.16113 0.01267
12 Hungary HU 129.000 446.044 271.176 16.0990 0.22717 0.11742
13 Ireland IE 86.348 211.119 129.577 18.6033 0.18991 0.02967
14 Italy IT 9 162.053 73.8 6.180 05 42.0046 0.12622 0.04835
15 Latvia LV 17 520.308 256.4 7.889 59 10.5700 0.38000 0.06000
16 Lithuania LT 223.625 498.111 190.550 8.9236 0.34000 0.05130
17 Luxembourg LU 76.168 115.759 54.111 21.1667 0.09677 0.01613
18 Malta MT 24.606 46.435 48.780 84.1000 0.35294 0.00000
19 The Netherlands NL 44.659 101.938 48.425 33.5986 0.09436 0.18288
20 Norway NO 52.200 116.000 62.848 32.7479 0.14463 0.03306
21 Poland PL 137.463 407.642 234.859 8.9407 0.32256 0.11076
22 Portugal PT 91.550 228.269 130.769 36.8215 0.15996 0.04128
23 Slovakia SK 107.393 439.060 214.986 13.7962 0.09300 0.05500
24 Slovenia SI 130.552 270.037 112.659 44.3511 0.13688 0.05323
25 Spain ES 93.027 200.748 117.211 24.3126 0.15308 0.01846
26 Sweden SE 49.006 106.509 44.724 40.7146 0.12360 0.05843
27 Switzerland CH 49.600 95.000 42.131 58.0946 0.20541 0.09459
127

Obs B7 C1 C2 C3 C4 C5 C6 C7
1 0.18356 0.05990 00 52.0000 20.5429 7.5148 0.12353 83.00 0.06974
29 08 00 23.8853 21.3000 0 0.155 0.044 71.00 6.650 0.05975 0.11133 2
0.29070 0.22549 79.6000 05 36.28 3 20.6755 9.3018 0.07589 0.21735
4 0.10913 0.05521 72.0000 13.0000 20.7000 11.0000 0.15113 0.09018
5 0.14706 0.22961 85.0000 63.0000 20.6679 7.3150 0.06746 0.18655
42 02 00 00 26 12.7269 0.01895 0.041 0.284 74.00 30.00 21.66 0.13972 6
99 83 60 78.0000 20.4537 9.1340 0.10471 0.094 0.234 87.97 0.11638 7
8 0.23487 0.28808 97.0000 70.0000 20.7287 8.3166 0.06420 0.13726
9 0.17678 0.05074 96.0000 89.0000 20.1000 6.0000 0.10466 0.05077
10 0.17506 95 00 00 00 0 0.174 90.00 84.00 20.60 5.000 0.03731 0.10656
94 76 06 34 00 0 0.299 0.106 79.41 28.02 20.70 8.000 0.17236 0.17438 11
54 64 00 34.0000 20.6000 0 0.100 0.087 67.00 8.000 0.03671 0.12531 12
20 37 00 46.0000 20.6434 8 0.163 0.149 86.00 5.135 0.01613 0.14726 13
16 0.02366 71.0000 61 38 1 0.238 21.29 20.95 7.856 0.20493 14 0.08670
15 00 19 00 90 00 0.040 0.217 77.00 22.51 21.30 10.0000 0.03725 0.12225
85 42 00 48 50 0.115 0.118 60.00 13.03 22.63 14.4799 0.01442 0.07663 16
0.00000 0.00000 00 00 20.7793 4.4223 0.10008 80.00 60.00 0.07960 17
47 80 00 00 00 0 0.04441 0.16429 0.176 0.249 96.30 28.00 18.60 9.000 18
85 33 00 00 00 0 0.183 0.153 90.00 64.00 20.20 6.000 0.06450 0.11309 19
20 0.15289 37 00 83.0000 21.0000 8.0000 0.308 91.00 0.09362 0.17036
21 0.03857 0.09827 77.7000 81 20.5527 11.7991 0.04711 0.14375 23.23
0.24149 0.11424 86.0000 45.0000 20.7379 7 9.674 0.09035 22 0.21346
0.07400 0.11964 76.2848 24.3845 20.9679 2 7.000 0.03654 23 0.11904
24 0.20532 0.32171 86.8500 30.0000 21.1468 4.9922 0.04812 0.06341
25 0.17650 0.12129 74.3500 51.0000 20.6213 7.6014 0.14563 0.16712
26 0.15730 0.34000 92.0000 73.0000 22.6000 7.0000 0.09584 0.09239
27 0.21622 0.19315 82.0000 53.0000 20.5469 7.2542 0.12499 0.06449

128

Obs N_A1 N_A2 N_A3 N_A4 N_A5 D1 D2
1 1.00000 3.00000 2.00000 2.00000 1.00000 506.689 98.92
2 2.00000 2.00000 2.00000 2.00000 2.00000 470.147 347.05
3 1.39650 2.27384 1.92987 1.75805 1.37147 478.943 83.76
4 2.00000 3.00000 2.00000 2.00000 2.00000 399.391 130.38
5 1.00000 2.00000 2.00000 2.00000 2.00000 370.843 126.38
6 2.00000 3.00000 2.00000 2.00000 1.00000 412.700 29.69
7 1.00000 2.00000 2.00000 2.00000 2.00000 474.808 15.61
8 1.00000 1.00000 2.00000 1.00000 1.00000 503.789 113.12
9 1.00000 2.00000 2.00000 1.00000 2.00000 565.750 230.57
10 1.00000 2.00000 2.00000 1.00000 1.00000 471.095 249.30
11 1.00000 2.00000 2.00000 2.00000 2.00000 406.652 84.64
12 1.00000 3.00000 2.00000 2.00000 2.00000 293.432 108.24
13 2.00000 2.00000 2.00000 1.00000 2.00000 417.682 61.38
14 3.00000 4.00000 3.00000 3.00000 4.00000 596.931 196.25
15 2.00000 3.00000 2.00000 2.00000 2.00000 360.325 35.31
16 2.00000 3.00000 2.00000 2.00000 2.00000 470.397 51.92
17 1.37723 3.63131 2.43701 1.90550 2.00012 660.913 183.08
18 1.00000 3.00000 2.00000 2.00000 2.00000 534.938 1360.00
19 1.00000 1.00000 2.00000 1.00000 1.00000 441.997 394.17
20 2.00000 1.00000 1.00000 2.00000 2.00000 445.233 14.44
21 2.00000 3.00000 2.00000 2.00000 2.00000 351.057 121.92
22 2.00000 3.00000 2.00000 2.00000 2.00000 404.752 115.33
23 2.00000 3.00000 2.00000 2.00000 2.00000 247.282 110.53
24 2.00000 3.00000 2.00000 2.00000 2.00000 487.601 99.01
25 1.00000 3.00000 2.00000 2.00000 2.00000 464.011 87.90
26 1.00000 1.00000 1.00000 1.00000 1.00000 461.137 20.25
27 1.00000 1.00000 1.00000 2.00000 1.00000 519.384 181.82

129
Appendix 2. Detailed results of the five analyses
A2.1. PCA in which all variables are analysed together (PCA-all)
Using the PCA, Eigenvalue>1 served as a criterion for choosing the number of
factors. As a result, five factors were chosen providing 76.3% of cumulative
explained variance. Then we used orthogonal rotation of the five factors. Scores
were calculated for each country for each factor. Then the five scores of each
country were weighted according to the weights created by the 'explained by each
factor' variance divided by the sum of the five variances.

Table A2.1 provides the rotated factor pattern received in this analysis, which makes
it possible to see which variables (basic indicators) contributed most to each one of
the factors built. The behaviours of major variables (with coefficients over 0.5),
which compose the factor, enable interpreting the 'safety-desirable' behaviour of the
factor. For example, for basic indicators C5 (median age of passenger cars), B1-B3
(number of fatalities per population, vehicles, passenger kms travelled), B5 (share of
pedestrian fatalities): the lower the values, the better the safety situation of the
country. Similarly, for indicators C2-C3 (safety belt wearing rates in front and rear
seats): the higher the values, the better for safety. These two groups of indicators
make a major contribution to Factor 1 (see Table A2.1), with positive and negative
coefficients, accordingly. Therefore, a better safety situation overall is associated
with lower values of Factor 1.

In Table A2.1 arrows indicate 'safety-desirable' behaviour of each basic indicator.
Based on this information and each factor's composition, the factor interpretations
will be as follows:
Factor 1 mainly reflects the safety 'product', car fleet's age and seatbelt use, and
for better safety it should aim at a lower value. Therefore, we can write: Factor 1
min;
Similarly, Factor 2 mainly reflects the 'strategy' indicators, but also includes a
negative correlation with C1 (share of drink-driving accidents). Therefore: Factor
2 min;
Factor 3 reflects the share of bicyclist fatalities, EuroNCAP scores for cars, and
population density. Therefore: Factor 3 min;
Factor 4 reflects the share of motorcycles in the vehicle fleet and the share of
motorcyclist fatalities. Therefore: Factor 4 min;
Factor 5 reflects the share of HGVs in the fleet, the number of injury accidents
per fatality, and the motorization level of the country, therefore the 'desirable'
value of this factor is unclear, but will probably be a minimum.

The tools produced by this analysis for the estimation of each country's score, i.e.
the factors' scoring coefficients, factors' weights, means and variances of the
variables to estimate the standardized values, are given in Appendix 3.

Table A2.2 provides the country scores estimated in this trial, where the combined
indicator is presented by WF (weighted index' value).

130
Basic indicator and its
'safety-desired'
behaviour Factor1 Factor2 Factor3 Factor4 Factor5
C5 0.82281 -0.05114 -0.13990 0.03258 0.07188
B2 0.81377 0.22390 -0.07361 -0.35261 0.27504
B1 0.80847 0.29893 -0.21667 -0.23356 0.04940
B3 0.74577 0.23969 0.00297 -0.30388 0.42035
B5 0.73546 -0.08223 -0.07773 -0.28072 0.06914
C2 -0.75692 -0.49112 -0.07181 0.00379 0.07170
C3 -0.77094 -0.47782 -0.01486 -0.01698 -0.06163
N_A3 -0.06187 0.86337 0.11585 -0.12262 0.07379
N_A2 0.32897 0.83821 -0.10312 -0.14772 -0.04356
N_A5 0.10671 0.78628 -0.10678 0.24352 0.01832
N_A4 0.48346 0.57833 -0.11082 0.40171 -0.02498
N_A1 0.38289 0.55035 -0.28421 -0.00448 -0.12514
C1 -0.01846 -0.70457 -0.48939 -0.07028 0.10070
B6 0.09152 -0.22500 0.83410 -0.09091 -0.08257
D2 -0.34390 0.16389 0.72886 0.03552 -0.23342
C4 0.49951 -0.17503 -0.54151 -0.17143 -0.41491
C6 -0.19535 0.27701 0.03949 0.81482 -0.13549
B7 -0.24746 -0.15919 -0.04603 0.72856 0.14163
C7 0.03620 -0.07892 -0.25342 0.19072 0.84957
B4 -0.48078 -0.17521 0.21314 0.33466 -0.52502
D1 -0.49118 0.18357 -0.16747 0.25852 -0.63870
Table A2.1. Rotated factor pattern received by the PCA in which all variables are
analysed together.


131
Obs Country Code Factor1 Factor2 Factor3 Factor4 Factor5 WF
1 Austria AT -0.18750 0.03685 0.50396 0.52196 -0.87198 -0.03163
2 Belgium BE 0.20785 0.49670 1.11582 0.14418 -0.88384 0.25906
3 Cyprus CY 0.20292 -0.36392 -0.69144 1.03772 1.26982 0.17758
4 Czech Republic CZ 1.06053 0.46159 1.03482 0.71011 -0.32119 0.68559
5 Denmark DK -0.45927 -0.33758 0.29466 0.11332 1.21670 -0.04707
6 Estonia EE 1.21917 -0.35021 -1.09404 -1.14521 -0.18884 0.02856
7 Finland FI -0.57484 -0.29836 -0.00991 0.09103 0.21920 -0.24463
8 France FR -1.10061 -0.96531 -0.76079 -0.25846 0.63887 -0.69691
9 Germany DE -1.41318 0.00544 1.09739 -0.24937 -1.01914 -0.51574
10 Great Britain UK -1.35379 -0.52159 0.31356 -0.96782 -0.39620 -0.74990
11 Greece EL 0.44094 0.19936 -0.32738 1.69338 1.55771 0.57649
12 Hungary HU 1.01982 0.22930 1.43842 -0.78585 1.07992 0.64505
13 Ireland IE -0.98623 0.29825 -0.79932 -1.25632 0.94042 -0.42463
14 Italy IT -0.11862 2.87298 -0.39783 2.03147 -1.21513 0.76132
15 Latvia LV 1.47182 0.11570 -0.46289 -1.30890 0.28562 0.36030
16 Lithuania LT 2.52497 -0.03205 -0.76131 -0.68984 -1.62858 0.50048
17 Luxembourg LU -1.71519 1.89584 -0.65024 -1.44137 -1.11003 -0.52814
18 Malta MT -1.10104 0.63562 4.22576 0.44586 -0.22985 0.35854
19 The Netherlands NL -0.82933 -0.90707 2.69329 -0.30286 0.13467 -0.19797
20 Norway NO -0.44158 -1.14757 -1.49153 0.89052 0.30156 -0.49961
21 Poland PL 1.28805 0.26680 0.94963 -0.81912 0.77487 0.64385
22 Portugal PT -0.12295 0.37701 -0.41458 0.95944 1.21851 0.27370
23 Slovakia SK 0.22268 0.72877 0.14052 -1.14595 1.00259 0.26355
24 Slovenia SI -0.22471 0.25272 -0.92202 -0.33291 -0.92573 -0.29356
25 Spain ES -0.31602 0.54563 -0.48487 0.83210 0.75748 0.16589
26 Sweden SE -0.22992 -2.19453 -1.35911 0.04294 -1.41079 -0.99941
27 Switzerland CH 0.41497 -1.66474 1.04518 1.63581 -1.42649 -0.11224
Table A2.2. Country scores estimated by the PCA in which all variables are
analysed together.

Based on the weighted index (WF) and using a WARD clustering procedure, the
countries can be classified into similar groups as presented in Figure A2.1. The
number of groups varies depending on the level of 'distances' between the
countries, which is selected as a threshold for countries in the same group. For
example, with a threshold value of 0.05, the countries are divided into four groups.

132
133

Figure A2.1. Similar groups of countries based on the PCA-all variables analysis.


Furthermore, we can recognize similarities in the behaviour of separate indicators
when they are plotted against the factors. For example, Figure A2.2 illustrates the
indicators' behaviour on the dimensions of Factor 1 and Factor 2, where similar
behaviour can be noted for variable groups of B1-B2-B3, A1-A4, A2-A3-A5, C2-C3.
The same groups of variables were relatively stable when they were plotted against
other factors, whereas other variables behaved differently on different dimensions.

Rotated Factor Pattern
-1
-0,8
-0,6
-0,4
-0,2
0
0,2
0,4
0,6
0,8
1
-1 -0,8 -0,6 -0,4 -0,2 0 0,2 0,4 0,6 0,8 1
Factor 1 (35.63%)
F
a
c
t
o
r

2

(
2
5
.
9
3
%
)
C3
C2
C1
C4
C5
C6
C7
B4
B2
B3
B5
B6
B7
N_A2
N_A4
N_A3
N_A5

N_A1
B1
D1
D2
Figure A2.2. Basic indicators plotted versus Factor 1 and Factor 2 (PCA-all
analysis).
PCA -all variables together
tree diagram Using METHOD=WARD, final weighted index
Index
Lithuania
Greece
Italy
Poland
Hungary
Czech Republic
Malta
Latvia
Spain
Cyprus
Portugal
Slovakia
Belgium
Ireland
Norway
Luxembourg
Germany
Sweden
Great Britain
France
Slovenia
Netherlands
Finland
Switzerland
Estonia
Denmark
Austria
Semi-Partial R-Squared
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
A2.2. PCA by groups of indicators (PCA-groups)
In this analysis we carried out the PCA for each group of indicators separately,
namely for the A, B, C and D groups of basic indicators. Initially, one combined
factor is fitted for each group. If more than one factor is obtained for a group, then
the factors are weighted to get one combined index (factor) for this group. The four
resulting combined factors are then subjected to another PCA, where the final
composite index is generated by using the weights attained from this second
analysis.

The eigenvalue>1 criterion was first used for choosing the number of factors for
each group of variables and next for the final combined indicator.

For the A-group indicators one factor was chosen which provided 64.42% of the
cumulative explained variance. According to the factor pattern received in this
analysis (Table A2.3a), each variable of the group contributes positively to the factor
value. As lower values of A1-A5 indicators are preferable (stating higher quality of
the national safety program), a lower value of the A-group factor is 'safety-desirable'.

The tools produced by the analysis of the A-group (and further, by the analyses of B,
C and D groups of indicators) for the estimation of each country's score, i.e. the
factors' scoring coefficients, factors' weights (if more than one factor received),
means, and variances of the variables (to estimate the standardized values), are
given in Appendix 3. Table A2.4 gives the country scores estimated for the A-group
(and for other groups) of indicators, where the countries' clustering into similar
groups (using the results of group analyses) is presented in Figure A2.3.

For the B-group indicators two factors were chosen providing 74.46% of cumulative
explained variance. According to the factor pattern received in this analysis (Table
A2.3b), variables B1-B2-B3-B5 contribute positively to the Factor 1 value, whereas
B4 (injury accidents per fatality) contributes negatively to the same factor; overall, a
lower value of Factor 1 can be considered as 'safety-desirable'. Two basic indicators
contribute to Factor 2: B6 (share of bicyclist fatalities) and B7 (share of motorcyclist
fatalities) but with opposite coefficients; overall, a lower value of Factor 2 can be
considered preferable for safety.

Exploring the C-group indicators (SPIs), low communalities with others were
observed for C6, C7 indicators (the percentages of motorcycles and HGVs in the
vehicle fleet, accordingly), which, consequently, were excluded from the analysis.
For the remaining indicators, two factors were chosen, providing 80.11% of
cumulative explained variance. According to the factor pattern obtained from this
analysis (Table A2.3c), variable C2-C3 (seatbelt wearing rate) contributes positively
to the Factor1 value, whereas C5 (median age of passenger cars) contributes
negatively to the same factor. Therefore, overall, a higher value of Factor 1 is
'safety-desirable'. Two basic indicators contribute to Factor 2: C1 (share of drink-
driving accidents) and C4 (average EuroNCAP score). However, as both indicators
contribute positively to Factor 2 (whereas minimum C1 and maximum C4 values are
'safety-desirable'), a preferred safety value cannot be given for Factor 2. In total,
having weighted both factors, a higher value of the C-group factor can be
considered safety-desirable.

For the D-group indicators (background characteristics), one factor was chosen
providing 63.2% of the cumulative explained variance. According to the factor
134
pattern received in this analysis (Table A2.3d), each variable of the group
contributes positively to the factor value. According to different sources, a higher
motorization level of the country and a lower population density are probably more
favourable for safety, therefore, the combined D-group factor appears to be safety-
indifferent.

As mentioned above, the tools produced by each group analysis for the estimation
of each country's score are given in Appendix 3, and the country scores estimated
for each group of indicators and the results of the countries' clustering into similar
groups are presented in Table A2.4 and Figure A2.3, respectively.

Indicators Factor1
N_A2 0.86565
N_A5 0.85723
N_A4 0.78888
N_A1 0.75598
N_A3 0.73698
a)

Indicators Factor1 Factor2
B2 0.96666 0.04130
B1 0.92452 -0.15173
B3 0.92295 0.04515
B5 0.76317 0.11852
B4 -0.73632 -0.14349
B6 -0.19054 0.84533
B7 -0.41192 -0.68164
b)

Indicators Factor1 Factor2
C2 0.95273 0.03179
C3 0.90504 -0.02923
C5 -0.64819 0.47955
C1 0.48383 0.80142
C4 -0.36733 0.78461
c)

Indicators Factor1
D2 0.79499
D1 0.79499
d)
Table A2.3. Factor patterns obtained from the PCA of different groups of indicators.

135
Obs
Country
Code
FA
(for A-
group)
FB
(final for
B-group)
FC
(final for
C-group)
FD
(for D-
group)
Factor1
(combined
analysis)
Factor2
(combined
analysis)
WF
(combined
index)
1 AT -0.21719 -0.53442 -0.32169 0.22792 0.06605 0.48602 0.25715
2 BE 0.30616 -0.25555 -0.85012 1.59920 1.01047 1.29614 1.14046
3 CY -0.28273 -0.19814 0.02848 -0.07009 -0.20786 0.05011 -0.09048
4 CZ 0.61473 0.40384 -1.08416 -0.33203 1.06280 -0.31133 0.43752
5 DK -0.10865 -0.17253 0.50540 -0.56240 -0.54665 -0.31531 -0.44138
6 EE 0.19761 1.05701 0.16487 -0.89934 -0.00478 -1.22661 -0.56076
7 FI -0.10865 -0.20836 0.69176 -0.54804 -0.67990 -0.30154 -0.50774
8 FR -1.33540 -0.60246 1.14367 0.30066 -1.49940 0.34913 -0.65826
9 DE -0.60972 -0.83857 0.33765 1.51628 -0.38600 1.41396 0.43304
10 UK -1.02683 -0.79988 0.68791 0.96286 -0.93618 0.96575 -0.07074
11 EL -0.10865 0.15535 -0.47386 -0.58107 0.18987 -0.43585 -0.09485
12 HU 0.19992 1.21408 -0.87098 -1.23510 0.67905 -1.44325 -0.28666
13 IE -0.19491 -0.04414 0.09534 -0.65524 -0.29888 -0.42234 -0.35506
14 IT 3.23882 -0.69045 -1.04289 1.51336 2.48709 1.75650 2.15465
15 LV 0.61473 1.80568 -0.10091 -1.23674 0.45282 -1.84347 -0.59206
16 LT 0.61473 1.50049 -0.85989 -0.34069 1.10238 -1.00584 0.14307
17 LU 0.75099 -0.25851 -0.45167 1.88401 1.00693 1.48082 1.22257
18 MT 0.19992 -1.03913 0.10246 8.72656 1.55859 6.29741 3.71491
19 NL -1.33540 -0.54444 0.30569 1.70799 -0.62884 1.30439 0.25084
20 NO -0.56783 -0.63929 1.23492 -0.76714 -1.40353 -0.26825 -0.88694
21 PL 0.61473 1.48981 -0.69580 -0.73316 0.90901 -1.26910 -0.08210
22 PT 0.61473 -0.38430 -0.09778 -0.39272 0.22289 0.03745 0.13850
23 SK 0.61473 0.65575 -0.48412 -1.54988 0.45084 -1.31034 -0.35055
24 SI 0.61473 -0.24852 0.70083 0.09214 -0.21273 0.19466 -0.02735
25 ES 0.19992 -0.25197 -0.34574 -0.14963 0.26371 0.10458 0.19130
26 SE -1.90082 -0.77338 1.62685 -0.61521 -2.32385 -0.23058 -1.37135
27 CH -1.39975 -0.83708 0.15625 0.86407 -0.77530 0.94428 0.00716
Table A2.4. Country scores resulting from the PCA for groups of indicators.


136
137

a) based on the A-group of indicators.


b) based on the B-group of indicators.
Figure A2.3. Similar groups of countries based on the 'PCA-by groups' analysis.

PCA -group A (1 factor chosen)
tree diagram
Index
Italy
Sweden
Great Britain
Switzerland
Netherlands
France
Luxembourg
Slovenia
Slovakia
Portugal
Poland
Lithuania
Latvia
Czech Republic
Norway
Germany
Spain
Malta
Hungary
Estonia
Belgium
Greece
Finland
Denmark
Cyprus
Ireland
Austria
Average Distance Between Clusters
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50

PCA - group B (2 factors chosen)
tree diagram, weighted index

Index

Poland

Lithuania

Latvia
Hungary
Estonia

Slovakia
Czech Republic

Ireland

Greece

Portugal

Denmark

Finland

Cyprus

Spain

Slovenia

Luxembourg
Belgium

Malta

Swe den
Great Britain

Switzerland

Germany

Italy

Norway

France

Netherlands

Austria

Average Distance Between Clusters

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.4

138


c) based on the C-group of indicators.


d) based on the D-group of indicators.
Figure A2.3. (cont.) Similar groups of countries based on the 'PCA-by groups'
analysis.


PCA -group C, no c7 c6 (2 factors chosen)
tree diagram, weighted index
Index
Sweden
Norway
France
Italy
Czech Republic
Poland
Hungary
Lithuania
Belgium
Slovenia
Great Britain
Finland
Netherlands
Germany
Denmark
Portugal
Latvia
Malta
Ireland
Switzerland
Estonia
Cyprus
Luxembourg
Slovakia
Greece
Spain
Austria
Average Distance Between Clusters
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.8
PCA -group D (1 factor chosen)
tree diagram
Index
Malta
Slovakia
Latvia
Hungary
Poland
Norway
Estonia
Sweden
Ireland
Greece
Finland
Denmark
Portugal
Lithuania
Czech Republic
Spain
Cyprus
Luxembourg
Netherlands
Italy
Germany
Belgium
Switzerland
Great Britain
Slovenia
France
Austria
Average Distance Between Clusters
0.00 0.25 0.50 0.75 1.0
0
1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25
139
At the second step of the analysis, the factors obtained for each group (i.e. FA, FB,
FC, FD in Table A2.4) were subjected to another PCA, to provide the final
composite index. At this step, two factors were chosen, providing 86.27% of
cumulative explained variance.

According to the factor pattern obtained from this analysis (Table A2.5), factors FA
and FC compose the Factor 1, where FD and FB compose Factor 2. Discussion of
previous interpretations of components of the group factors, has shown that it is
safer to have lower values of FA and FB factors and a higher value of the FC factor.
Therefore, we can state that a lower value of the combined Factor 1 is 'safety-
desirable', whereas for the combined Factor 2 we should probably prefer a higher
value.

The tools produced by the estimation of the final safety indicator in this case, i.e. the
factors' scoring coefficients, factors' weights, means and variances, are discussed in
Appendix 3. The last three columns of Table A2.4 contain the country scores
estimated for each factor of the combined analysis and the final composite index,
whereas the countries' clustering into similar groups, using the final indicator, is
presented in Figure A2.4.

Rotated Factor Pattern
Indicator (group factor) Factor1 Factor2
FA 0.89253 -0.05656
FC -0.92977 0.08105
FD 0.13049 0.94526
FB 0.45004 -0.81670
Table A2.5. Factor patterns received by the PCA for four indices (FA, FB, FC, FD).


Figure A2.4. Countries' clustering into similar groups, using the final indicator of the
PCA-by groups' analysis.
PCA -group ABCD (2 factors chosen)
tree diagram Using METHOD=WARD, final weighted index
Index
Malta
Italy
Luxembourg
Belgium
Sweden
Norway
Slovakia
Ireland
Hungary
France
Latvia
Estonia
Finland
Denmark
Switzerland
Slovenia
Poland
Great Britain
Greece
Cyprus
Germany
Czech Republic
Spain
Portugal
Lithuania
Netherlands
Austria
Semi-Partial R-Squared
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70
A2.3. Common factor analysis with four factors (FA-4Factors)
General comment

A preliminary examination of all variables together revealed a low value of the
overall KMO statistics (0.45), and the lowest values of the KMO statistics for such
variables as B6 (0.131), B7 (0.309), and C7 (0.136). Once the indicators B6, B7,
and C7 were excluded, the overall KMO statistics increased to 0.63, making it
possible to proceed with the Common Factor Analysis (FA). Therefore, the coming
three trials of creating the composite index by means of FA consider 18 variables
(basic indicators), without B6 (share of bicyclist fatalities), B7 (share of motorcyclist
fatalities), and C7 (percentage of HGVs in fleet).

To determine the number of factors required, the variance explained criterion is
applied. At this point, two approaches are possible. Some researchers use the rule
of keeping enough factors to account for certain share (e.g. 90% or 80%) of the
variation. Conversely, if the researcher's goal emphasizes parsimony (explaining
variance with as few factors as possible), the criterion could be as low as 50%. In
our case, the variation was taken to be the sum of the eigenvalues of the reduced
correlation matrix. It was found that four factors explain more than 80% of the
variation.

Furthermore, examining the solution for fou demonstrated that C4 variable
(average EuroNCAP score) is nearly identical to one of the factors. Moreover,
excluding C4 resulted in two factors being responsible for 71% of the variation.
Hence, we decided to explore the results of three trials:

FA-4Factors FA with all variables (except for B6, B7, C7 as explained
before) and four factors.
FA-2Factors-noC4 FA with the same variables excluding C4 and two factors.
FA-2factors FA with the above variables, including C4, and two factors.

The FA-2Factors results can be conveniently displayed, and give insight into the
data.

FA-4Factors

In this case, the average of the eigenvalues of the reduced correlation matrix was
0.87. Four factors explained 84% of cumulative variance.

Table A2.6 provides the rotated factor pattern obtained from this analysis, which
shows which variables (basic indicators) contributed more to each one of the factors
built. The behaviours of major variables (with coefficients over 0.5), which compose
the factor, make it possible to interpret the 'safety-desirable' behaviour of the factor.

As discussed previously, for basic indicators B1-B3 (number of fatalities per
population, vehicles, passenger kms travelled), B5 (share of pedestrian fatalities),
C5 (median age of passenger cars), lower values are safety-preferable. Similarly, for
indicators C2-C3 (safety belt wearing rates in front and rear seats), higher values
are better for safety. Concerning B4 (injury accidents per fatality), a lower value is
safety-desirable, whereas for D1 (number of passenger cars per population) a
higher value is probably associated with better safety. All these indicators make a
major contribution to Factor 1 (see Table A2.6). Accounting for the basic indicators'
r factors
140
effect on Factor 1's value (positive or negative coefficients), we can conclude that a
better safety situation is associated with a lower value of Factor 1.

Factor 2 (see Table A2.6) mainly reflects the behaviour of 'safety strategy' indicators
A1, A2, A4, A5 and C6 (percentage of motorcycles in the vehicle fleet), for which
lower values are safety-preferable. Hence, lower value of Factor 2 should be seen
as safety-preferable.

Factor 3 consists of the C4 indicator only (average EuroNCAP score) which should
be higher from a safety point of view.

Factor 4 reflects the behaviour of D2 (population density), A3 (quality of economic
basis of safety program) and C1 (share of drink-driving accidents). For increased
safety, all three variables should have lower values. Accounting for their coefficients
in the Factor 4 composition, the safety-desirable behaviour of Factor 4 is not
obvious, but a lower value is probably preferable.

To sum up the factors' interpretation and characterize the safety-desirable factors'
behaviour, we can give the following summary:
Factor 1 min; Factor 2 min; Factor 3 max; Factor 4 min (not obvious).

Rotated Factor Pattern
Basic indicator
and its 'safety-
desired' behaviour Factor1 Factor2 Factor3 Factor4
B3 0.93114 0.16237 0.16430 0.09886
B2 0.90113 0.15766 0.34291 0.07174
B1 0.68637 0.24702 0.59423 0.11426
B5 0.60411 -0.03942 0.45413 -0.06381
C5 0.52698 0.20849 0.46343 -0.23475
C2 -0.50146 -0.49013 -0.43328 -0.28767
C3 -0.58789 -0.47246 -0.41165 -0.25516
B4 -0.70392 -0.17008 -0.09415 0.14029
D1 -0.82127 0.13857 0.19451 0.17435
N_A4 0.18750 0.85028 0.13938 -0.01902
N_A5 0.02815 0.81876 -0.00825 0.19648
N_A2 0.31802 0.65628 0.22076 0.41407
C6 -0.45550 0.58664 -0.30691 -0.05497
N_A1 0.15967 0.55971 0.44619 0.10842
C4 0.09648 -0.00215 0.77306 -0.32513
D2 -0.36120 -0.19337 -0.16500 0.71500
N_A3 0.09619 0.48215 -0.01314 0.65940
C1 -0.02361 -0.40141 0.13150 -0.73881
Table A2.6. Rotated factor pattern for FA-4Factors' trial.


141
The tools obtained from this analysis for the estimation of each country's score, i.e.
the factors' scoring coefficients, factors' weights, means, and variances of the
variables to estimate the standardized values, are given in Appendix 4.

Table A2.7 provides the country scores estimated in this trial, where the composite
index is presented by WF-4F (weighted index' value).

Obs Country Code Factor1 Factor2 Factor3 Factor4
WF-4F
(combined
index)
1 Austria AT -0.51333 0.11212 -0.36543 0.30850 -0.18372
2 Belgium BE -0.61693 -0.10196 0.86891 1.48278 0.12586
3 Cyprus CY 0.28024 -0.16952 0.05955 -0.19428 0.04165
4 Czech Republic CZ 0.32629 0.98406 -0.00322 0.25801 0.43665
5 Denmark DK -0.11726 0.11485 -1.07465 -0.67545 -0.31372
6 Estonia EE 0.92728 -0.34931 1.61377 -0.53938 0.46403
7 Finland FI -0.32609 0.52476 -1.06877 -1.25634 -0.37508
8 France FR -0.27240 -1.23411 -0.44444 -0.53126 -0.60764
9 Germany DE -1.03623 -0.74971 -0.72937 1.06208 -0.56560
10 Great Britain UK -0.76593 -1.72859 -0.02897 1.00974 -0.61245
11 Greece EL 0.81396 0.60818 -0.73238 -0.31280 0.30160
12 Hungary HU 2.08833 -0.08571 -1.10172 0.59256 0.68627
13 Ireland IE 0.17444 -0.64890 -0.16791 0.26594 -0.09682
14 Italy IT -1.30599 3.26270 0.21343 0.90846 0.56986
15 Latvia LV 1.77916 -0.00079 0.64996 -0.26562 0.76263
16 Lithuania LT 0.79314 0.01150 3.34107 0.18876 0.93474
17 Luxembourg LU -1.33076 0.61825 0.11757 1.32288 -0.11329
18 Malta MT -1.80332 -1.89677 0.01888 5.67029 -0.30215
19 The Netherlands NL -0.54731 -1.69909 -0.76365 1.34816 -0.59562
20 Norway NO -0.81465 0.19592 -0.11620 -2.37984 -0.66613
21 Poland PL 1.49531 0.21653 -0.19155 0.27671 0.64975
22 Portugal PT -0.05081 0.56496 -0.27798 -0.10442 0.06860
23 Slovakia SK 1.73187 0.15183 -0.95102 0.26927 0.58770
24 Slovenia SI -0.48375 0.20160 1.05057 -0.15750 0.02878
25 Spain ES -0.16303 0.72902 -0.40731 -0.08646 0.04986
26 Sweden SE -1.09173 -1.08386 0.61912 -2.09179 -0.94661
27 Switzerland CH -0.97385 -0.44472 -0.10937 -0.69870 -0.63127
Table A2.7. Country scores estimated by FA with four factors' solution.


Based on the final weighted index (WF-4F) and using a WARD clustering procedure,
the countries can be classified into similar groups, as presented in Figure A2.5a (the
countries' clustering based on the results of other two FA trials are presented in the
same figure). As usual, the number of groups can vary depending on the level of the
142
143
'distance' between the countries in the same group, which is selected as a
threshold. For example, with a threshold value of about 0.02, five groups of
countries can be distinguished.


a) FA with four factors' solution.


b) FA with two factors' solution (C4 excluded).
Figure A2.5. Countries' clustering into similar groups based on FA analyses.



FA 4 Factors' solution
tree diagram Using METHOD=WARD, final weighted index
Index

Lithuania

Slovakia
Italy
Latvia

Poland

Hungary
Greece

Estonia

Czech Republic

Sweden
Norway
Germany

Switzerland

Netherlands

Great Britain

France

Portugal

Slovenia
Spain

Cyprus

Belgium

Finland

Malta

Denmark

Luxembourg

Ireland

Austria

Semi-Partial R-Squared

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8


FA 2

Factors' solution no C4
tree diagram Using METHOD=WARD, final weighted index

Index

Lithuania

Latvia

Poland

Hungary
Slovakia
Estonia

Greece

Italy
Czech Republic

Sweden
Netherlands

Great Britain

Germany

Norway
Switzerland

Malta

France

Portugal

Slovenia
Cyprus

Spain

Be lgium
Finland

Denmark

Ireland

Luxembourg

Austria

Semi-Partial R-Squared

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

144

c) FA with two factors' solution (C4 included).
Figure A2.5. (cont.) Countries' clustering into similar groups based on FA analyses.
A2.4. Common factor analysis with two factors, C4 excluded (FA-
2Factors-noC4)
In this analysis, the average of the eigenvalues of the reduced correlation matrix
was 0.85. Two factors explained 71.8% of cumulative variance.

Table A2.8 provides the rotated factor pattern received in this analysis. It can be
seen that Factor 1 mainly reflects the behaviour of B1-B5 indicators ('safety
product'), D1-D2 ('background characteristics'), C5 (median age of passenger cars),
and C6 (percentage of motorcycles in the vehicle fleet). For all these variables,
except for D1, lower values are safety-preferable. Accounting for the indicators'
correlations with the Factor 1 value (see Table A2.8), a lower value of this factor can
be considered as safety-preferable (although, an 'opposite contribution' to Factor 1
or a 'moderating effect' is expected from the B4, C6, and D2 indicators).

Factor 2 (see Table A2.8) mainly reflects the behaviour of 'safety strategy' indicators
A1-A5, C2-C3 (seatbelt use) and C1 (share of drink-driving accidents). For A1-A5
and C1 lower values are safety-preferable, whereas for C2-C3 higher values are
desirable. Hence, a lower value of Factor 2 should be considered safety-preferable
(with a 'moderating effect' expected from C1 indicator).

To sum up the factors' interpretation and to characterize the safety-desirable factors'
behaviour, we can write:
Factor 1 min; Factor 2 min.


FA 2 Factors' solution with C4
tree diagram Using METHOD=WARD, final weighted index

Index

Lithuania

Latvia

Slovakia
Poland

Hungary
Estonia

Greece

Italy
Czech Republic

Malta

Sweden
N etherlands

Great Britain

Germany

Norway
Switzerland

France

Slovenia
Portugal

Cyprus

Spain

Belgium

Ireland

Finland

Denmark

Luxembourg

Austria

Semi-Partial R-Squared

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Rotated Factor Pattern
Basic indicator and its' safety-
desirable' behaviour Factor1 Factor2
B2 0.91349 0.31392
B3 0.86582 0.30533
B1 0.81085 0.42604
B5 0.74210 0.05929
C5 0.70113 0.17389
D2 -0.53665 0.14608
C6 -0.53827 0.36909
B4 -0.68467 -0.15757
D1 -0.69604 0.15062
N_A2 0.25093 0.82875
N_A5 -0.06048 0.79124
N_A3 -0.08222 0.74827
N_A4 0.19439 0.73048
N_A1 0.26819 0.61014
C3 -0.60687 -0.65187
C2 -0.52413 -0.67475
C1 0.20785 -0.69421
Table A2.8. Rotated factor pattern for FA 2factors' trial (C4 excluded).


The tools obtained from this analysis for the estimation of each country's score, i.e.
the factors' scoring coefficients, factors' weights, means and variances of the
variables to estimate the standardized values, are given in Appendix 4.

Table A2.9 provides the country scores estimated in this trial, where the composite
index is presented by WF (weighted index' value).

Similarities and differences in the behaviours of separate indicators can be
considered when they are plotted against the factors (Figure A2.6). As Figure A2.6
illustrates, groups of indicators with similar behaviour are: B1-B2-B3-B5-C5 ('road
safety performance indicator' and 'median age of cars'), A1-A2-A3-A4-A5 ('policy
performance indicator'), C2-C3 (use of seatbelts), D1-D2-C6 ('background charac-
teristics' and percentage of motorcycles in fleet), where C1 (share of drink-driving
accidents) and B4 (injury accidents per fatality) demonstrate 'individual' behaviour.


145
Obs Country Code Factor1 Factor2 WF
1 Austria AT -0.58748 0.06590 -0.28994
2 Belgium BE -0.53285 0.75390 0.05312
3 Cyprus CY 0.35085 -0.23072 0.08601
4 Czech Republic CZ 0.15384 1.01263 0.54492
5 Denmark DK -0.35010 -0.42282 -0.38322
6 Estonia EE 1.58534 -0.26214 0.74402
7 Finland FI -0.44325 -0.45957 -0.45068
8 France FR -0.22973 -1.40723 -0.76595
9 Germany DE -1.40834 -0.27982 -0.89442
10 Great Britain UK -0.84980 -1.02428 -0.92926
11 Greece EL 0.50645 0.30152 0.41313
12 Hungary HU 1.25743 0.28024 0.81243
13 Ireland IE 0.05364 -0.44329 -0.17266
14 Italy IT -1.44082 3.11863 0.63550
15 Latvia LV 1.92435 0.15291 1.11766
16 Lithuania LT 1.95693 0.71593 1.39180
17 Luxembourg LU -1.46607 1.10102 -0.29704
18 Malta MT -2.80077 1.68566 -0.75770
19 Netherlands NL -1.02611 -0.88366 -0.96124
20 Norway NO -0.24445 -1.22666 -0.69174
21 Poland PL 1.25358 0.44864 0.88702
22 Portugal PT -0.09297 0.37697 0.12104
23 Slovakia SK 1.03545 0.31605 0.70784
24 Slovenia SI -0.00394 0.22419 0.09995
25 Spain ES -0.37206 0.49854 0.02440
26 Sweden SE -0.30211 -1.81707 -0.99200
27 Switzerland CH -0.72780 -0.90981 -0.81068
Table A2.9. Country scores estimated by FA with two factors' solution (C4 excluded).


146
Rotated Factor Pattern
-1
-0,8
-0,6
-0,4
-0,2
0
0,2
0,4
0,6
0,8
1
-1 -0,8 -0,6 -0,4 -0,2 0 0,2 0,4 0,6 0,8 1
Factor 1 (54.46%)
F
a
c
t
o
r

2

(
4
5
.
5
4
%
)
D2
D1
B4
C3
C2
C1
B5
C5
B2
B3
B1
N_A1
N_A2
N_A5
N_A4 N_A3
C6

Figure A2.6. Basic indicators plotted versus factors, according to FA with two
factors' solution (C4 excluded).


Based on the final weighted index (WF) of this trial, the countries can be classified
into similar groups, as presented in Figure A2.5b. For example, with a threshold
value of 0.025, five groups of countries can be recognized. Figure A2.7 presents the
countries' subdivision into five groups, where the country positions are plotted using
the weighted index (WF) and Factor 1 values. It can be seen that: the countries with
the best safety level (first group) are Sweden, the Netherlands, Great Britain,
Germany, Switzerland, France, Malta and Norway. The second group includes
Finland, Denmark, Austria, Luxembourg and Ireland. The third group consists of
Spain, Belgium, Cyprus, Slovenia and Portugal. The fourth group consists of
Greece, Czech Republic, Italy, Hungary, Estonia, Slovakia and Poland, and the fifth
group consists of Latvia and Lithuania.

Figure A2.7 shows that Malta, Italy and Luxembourg (to a less extent) behave as
'outsiders' of their groups (similar to the results of FA with four factors see Section
A2.3). These three countries are characterized by much lower values of Factor1 (in
general, and especially in comparison with the countries of their groups, i.e. with
similar level of the composite index) but by high positive values of Factor 2.

147
FA - all variables together
Countries-WF*Factor1(out of 2), using weighted factor for clustering
-3
-2
-1
0
1
2
-1 -0,5 0 0,5 1 1,5
WF
F
a
c
t
o
r

1
1 2 3 4 5
Lithuania Latvia
Poland
Czech Republic
Hungary
Slovakia
Belgium
Greece
Portugal
Estonia
Spain
Italy
Malta
Cyprus
Slovenia
Switzerland
Austria
Ireland
Denmark
Finland
Norway
France
Sweden
Luxembourg
Great Britain
Netherlands
Germany
Cluster membership

Figure A2.7. Countries plotted using WF values and Factor 1 values (FA with two
factors', C4 excluded).
A2.5 Common factor analysis with two factors, C4 included (FA-
2Factors)
In this analysis the average of the eigenvalues of the reduced correlation matrix was
0.87. Two factors explained 68.7% of cumulative variance.

Table A2.10 provides the rotated factor pattern received in this analysis. It is clear
that Factor 1 mainly reflects the behaviour of B1-B5 indicators ('safety performance
indicator final outcome'), D1-D2 ('background characteristics'), C4 (average
EuroNCAP score), C5 (median age of passenger cars) and C6 (percentage of
motorcycles in vehicle fleet). For the majority of these variables, except for D1 and
C4, lower values are safety-preferable. Accounting for the indicators' correlations
with Factor 1 (see Table A2.10), lower value of this factor can be considered as
safety-preferable (although, an 'opposite contribution' to Factor 1 or a 'moderating
effect' is expected from C4, C6, D2, and B4 indicators).

Factor 2 (see Table A2.10) reflects mostly the behaviour of 'policy performance
indicators' A1-A5, C2-C3 (safety belt use) and C1 (share of drink-driving accidents)
indicators. For A1-A5 and C1 lower values are safety-preferable, whereas for C2-C3
higher values are desired. Hence, lower value of Factor 2 should be seen as safety-
preferable (with a 'moderating effect' expected from the C1 indicator).

We can summarize the factors' interpretation and characterize the safety-desirable
factors' behaviour as follows: Factor 1 min; Factor 2 min. In general, this FA
solution is similar to that received in the previous trial (see Section A2.4).

148
Rotated Factor Pattern
Basic indicator and
its 'safety-desired
behaviour' Factor1 Factor2
B2 0.89030 0.35273
B3 0.82662 0.35248
B1 0.80647 0.44793
B5 0.73794 0.08673
C5 0.70792 0.19072
C4 0.52185 -0.05787
C6 -0.54598 0.34493
D2 -0.56129 0.13888
D1 -0.66133 0.10334
B4 -0.66607 -0.18853
N_A2 0.21927 0.83801
N_A5 -0.08187 0.78293
N_A3 -0.13059 0.75847
N_A4 0.17631 0.73949
N_A1 0.27601 0.60189
C3 -0.58845 -0.66931
C2 -0.51539 -0.69034
C1 0.26249 -0.70695
Table A2.10. Rotated factor pattern for FA 2factors' trial (C4 included).


The tools produced by this analysis for the estimation of each country's score, i.e.
the factors' scoring coefficients, factors' weights, means and variances of the
variables to estimate the standardized values, are given in Appendix 4.

Table A2.11 provides the country scores estimated in this trial, where the composite
index is presented by WF (weighted index' value).

Based on the final weighted index (WF) of this trial, the countries can be classified
into similar groups, as presented in Figure A2.5c. For example, with a threshold
value of 0.03, the same five groups of countries can be recognized as in Figure
A2.7.

149
Obs Country Code Factor1 Factor2
WF
(combined index)
1 Austria AT -0.69070 0.12768 -0.31567
2 Belgium BE -0.51748 0.71218 0.04603
3 Cyprus CY 0.32665 -0.19086 0.08949
4 Czech Republic CZ 0.14626 0.99895 0.53702
5 Denmark DK -0.41520 -0.41776 -0.41637
6 Estonia EE 1.68690 -0.21917 0.81341
7 Finland FI -0.45790 -0.41664 -0.43899
8 France FR -0.21657 -1.39872 -0.75831
9 Germany DE -1.44982 -0.31614 -0.93030
10 Great Britain UK -0.82861 -1.01949 -0.91608
11 Greece EL 0.40200 0.36988 0.38728
12 Hungary HU 1.11044 0.42401 0.79588
13 Ireland IE 0.03856 -0.39053 -0.15808
14 Italy IT -1.44760 2.99680 0.58911
15 Latvia LV 1.87350 0.22088 1.11617
16 Lithuania LT 2.14449 0.66436 1.46620
17 Luxembourg LU -1.46473 1.01450 -0.32859
18 Malta MT -2.92575 1.13839 -1.06330
19 Netherlands NL -1.07151 -0.87796 -0.98281
20 Norway NO -0.11523 -1.23237 -0.62717
21 Poland PL 1.09063 0.53270 0.83495
22 Portugal PT -0.17575 0.36781 0.07334
23 Slovakia SK 0.95011 0.41402 0.70444
24 Slovenia SI 0.10026 0.14407 0.12033
25 Spain ES -0.34440 0.47855 0.03273
26 Sweden SE -0.01460 -2.08592 -0.96381
27 Switzerland CH -0.65969 -0.90085 -0.77020
Table A2.11. Country scores estimated by FA with two factors (C4 included).




150
Appendix 3. Tools produced by Principle
Component analyses for the estimation of country
scores
A3.1. PCA with all variables together

The FACTOR Procedure
Rotation Method: Varimax

Scoring coefficients estimated by regression

Standardized Scoring Coefficients
Factor1 Factor2 Factor3 Factor4 Factor5
C5 C5 0.25408 -0.13566 0.04206 0.19032 -0.05939
B2 B2 0.11246 0.00784 0.03175 -0.08294 0.06772
B1 B1 0.13104 0.01431 -0.04654 -0.03323 -0.05776
B3 B3 0.09164 0.02078 0.07020 -0.05871 0.16271
B5 B5 0.17343 -0.09751 0.04047 -0.02016 -0.05770
C2 C2 -0.18502 -0.02891 -0.10493 -0.10049 0.09878
C3 C3 -0.17300 -0.03140 -0.08181 -0.11654 0.02552
N_A3 -0.17408 0.29537 -0.01147 -0.18716 0.09944
N_A2 -0.06459 0.23706 -0.08035 -0.13818 -0.02020
N_A5 -0.05376 0.21097 -0.07532 0.07340 0.04606
N_A4 0.13841 0.06425 -0.00165 0.27847 -0.03116
N_A1 0.01334 0.12693 -0.14181 -0.01587 -0.09111
C1 C1 0.01272 -0.17374 -0.22246 0.00030 0.00650
B6 B6 0.16643 -0.13471 0.46687 0.05460 -0.05344
D2 D2 -0.00532 0.03961 0.33310 -0.01545 -0.06712
C4 C4 0.12877 -0.10057 -0.23687 -0.03588 -0.32222
C6 C6 0.08565 0.00763 0.03975 0.44775 -0.02143
B7 B7 0.07552 -0.09037 0.01519 0.43406 0.11479
C7 C7 -0.06784 0.01158 -0.09112 0.13671 0.47712
B4 B4 0.03136 -0.06600 0.07723 0.14156 -0.24407
D1 D1 -0.08177 0.07762 -0.16170 0.01309 -0.29788

Proportion of variance explained by each factor

Obs Factor1 Factor2 Factor3 Factor4 Factor5 k sumF
1 5.7106158 4.1566576 2.1129504 2.0627407 1.9850024 1 16.0280

151
Factors' weights

Obs w1 w2 w3 w4 w5
1 0.35629 0.25934 0.13183 0.12870 0.12385

Means and variances of variables

Obs _TYPE_ B1 B2 B3 B4 B5 B6
1 MEAN 98.5564 240.058 120.029 29.0867 0.18001 0.065286
2 STD 44.0645 135.296 71.095 16.9387 0.07688 0.040203

Obs B7 C1 C2 C3 C4 C5 C6 C7
1 0.15433 0.15846 81.2374 46.5639 20.9006 8.08366 0.082161 0.12250
2 0.07748 0.09822 9.1263 23.4846 0.6053 2.44772 0.049863 0.04579

Obs N_A1 N_A2 N_A3 N_A4 N_A5 D1 D2
1 1.52976 2.38097 1.93719 1.79475 1.78352 449.344 126.575
2 0.56578 0.87092 0.40464 0.48876 0.63802 87.989 95.598
A3.2. PCA with groups of indicators
A-group
Scoring coefficients estimated by regression

Standardized
Scoring
Coefficients
Factor1
N_A2 0.26874
N_A5 0.26612
N_A4 0.24491
N_A1 0.23469
N_A3 0.22879

Means and variances of variables

Obs _TYPE_ N_A1 N_A2 N_A3 N_A4 N_A5
1 MEAN 1.52976 2.38097 1.93719 1.79475 1.78352
2 STD 0.56578 0.87092 0.40464 0.48876 0.63802
152
B-group
Scoring coefficients estimated by regression

Standardized Scoring Coefficients
Factor1 Factor2
B2 B2 0.24452 -0.01739
B1 B1 0.24397 -0.17287
B3 B3 0.23316 -0.01193
B5 B5 0.18850 0.05646
B4 B4 -0.18033 -0.07828
B6 B6 -0.09334 0.70073
B7 B7 -0.06907 -0.53512

Proportion of variance explained by each factor

Obs Factor1 Factor2 k sumF w1 w2
1 3.9715847 1.2406149 1 5.21220 0.76198 0.23802

Means and variances of variables

Obs _TYPE_ B1 B2 B3 B4 B5 B6 B7
1 MEAN 98.5564 240.058 120.029 29.0867 0.18001 0.065286 0.15433
2 STD 44.0645 135.296 71.095 16.9387 0.07688 0.040203 0.07748
C-group
Scoring coefficients estimated by regression

Standardized Scoring
Coefficients
Factor1 Factor2
C2 C2 0.38485 0.07494
C3 C3 0.36226 0.03083
C5 C5 -0.23377 0.28935
C1 C1 0.23942 0.57131
C4 C4 -0.10376 0.51224

Proportion of variance explained by each factor

Obs Factor1 Factor2 k sumF w1 w2
1 2.5159608 1.4897073 1 4.00567 0.62810 0.37190

153
Means and variances of variables

Obs _TYPE_ C1 C2 C3 C4 C5
1 MEAN 0.15846 81.2374 46.5639 20.9006 8.08366
2 STD 0.09822 9.1263 23.4846 0.6053 2.44772
D-group
Scoring coefficients estimated by regression

Standardized Scoring
Coefficients
Factor1
D2 D2 0.62894
D1 D1 0.62894

Means and variances of variables

Obs _TYPE_ D1 D2
1 MEAN 449.344 126.575
2 STD 87.989 95.598
Combined indicator
Scoring coefficients estimated by regression

Standardized Scoring
Coefficients
Factor1 Factor2
FA 0.49023 0.07950
FC -0.50778 -0.06804
FD 0.19696 0.64838
FB 0.14362 -0.48625

Proportion of variance explained by each factor

Obs Factor1 Factor2 k sumF w1 w2
1 1.8806542 1.5702955 1 3.45095 0.54497 0.45503

Means and variances of variables

Obs _TYPE_ FA FB FC FD
1 MEAN 0 0.00000 0.00000 0
2 STD 1 0.79829 0.72994 1


154
Appendix 4. Tools produced by common factor
analyses for the estimation of country scores
A4.1. FA with four factors' solution
Scoring coefficients estimated by regression

Standardized Scoring Coefficients
Factor1 Factor2 Factor3 Factor4
B3 B3 0.28284 0.03886 -0.47369 -0.04605
B2 B2 0.83187 -0.19123 -0.34775 0.20168
B1 B1 -0.43343 -0.07134 0.96965 0.03006
B5 B5 0.01655 -0.05018 0.05501 -0.01055
C5 C5 0.09078 0.05955 0.04620 -0.10329
C2 C2 -0.08465 0.04512 -0.33210 -0.20254
C3 C3 -0.06120 -0.16283 0.08725 -0.09027
B4 B4 -0.04864 -0.05139 0.04067 0.10191
D1 D1 -0.10739 0.02804 0.12406 0.13266
N_A4 -0.04051 0.37078 -0.23965 -0.28150
N_A5 -0.02689 0.32963 -0.13568 -0.05602
N_A2 -0.12188 0.14262 0.26336 0.16829
C6 C6 -0.06311 0.22161 -0.08883 -0.14550
N_A1 -0.08813 0.03893 0.34375 0.01278
C4 C4 -0.14834 0.00998 0.21873 -0.15999
D2 D2 -0.10885 -0.17951 0.14318 0.41278
N_A3 0.13289 0.04168 -0.28114 0.24323
C1 C1 0.03043 -0.04445 0.14917 -0.24055

Proportion of variance explained by each factor

Obs Factor1 Factor2 Factor3 Factor4 k sumF
1 5.0795326 3.5812942 2.3264115 2.1078804 1 13.0951

Obs w1 w2 W3 w4
1 0.38790 0.27348 0.17765 0.16097

Means and variances of variables

Obs _TYPE_ B1 B2 B3 B4 B5
1 MEAN 98.5564 240.058 120.029 29.0867 0.18001
2 STD 44.0645 135.296 71.095 16.9387 0.07688

155
Obs C1 C2 C3 C4 C5 C6
1 0.15846 81.2374 46.5639 20.9006 8.08366 0.082161
2 0.09822 9.1263 23.4846 0.6053 2.44772 0.049863

Obs N_A1 N_A2 N_A3 N_A4 N_A5 D1 D2
1 1.52976 2.38097 1.93719 1.79475 1.78352 449.344 126.575
2 0.56578 0.87092 0.40464 0.48876 0.63802 87.989 95.598
A4.2. FA with two factors' solution (C4 excluded)
Scoring coefficients estimated by regression

Standardized Scoring Coefficients
Factor1 Factor2
B2 B2 0.45090 -0.01603
B3 B3 0.11485 -0.00612
B1 B1 0.06441 0.00304
B5 B5 0.06424 -0.03844
C5 C5 0.11670 0.02916
D2 D2 -0.13370 0.11491
C6 C6 -0.09095 0.10344
B4 B4 -0.04894 -0.00145
D1 D1 -0.10158 0.10990
N_A2 -0.08341 0.29142
N_A5 -0.10388 0.27347
N_A3 -0.02490 0.10401
N_A4 -0.01382 0.08241
N_A1 0.04763 0.04050
C3 C3 -0.00070 -0.21626
C2 C2 -0.08587 -0.09107
C1 C1 0.08981 -0.11036

Proportion of variance explained by each factor

Obs Factor1 Factor2 k sumF w1 w2
1 5.6841616 4.7529446 1 10.4371 0.54461 0.45539

Means and variances of variables

Obs _TYPE_ B1 B2 B3 B4 B5
1 MEAN 98.5564 240.058 120.029 29.0867 0.18001
2 STD 44.0645 135.296 71.095 16.9387 0.07688

156
Obs C1 C2 C3 C5 C6
1 0.15846 81.2374 46.5639 8.08366 0.082161
2 0.09822 9.1263 23.4846 2.44772 0.049863

Obs N_A1 N_A2 N_A3 N_A4 N_A5 D1 D2
1 1.52976 2.38097 1.93719 1.79475 1.78352 449.344 126.575
2 0.56578 0.87092 0.40464 0.48876 0.63802 87.989 95.598
A4.3. FA with two factors' solution (C4 included)
Scoring coefficients estimated by regression

Standardized Scoring Coefficients
Factor1 Factor2
B2 B2 0.50258 0.03286
B3 B3 0.03320 -0.00930
B1 B1 0.07194 0.01471
B5 B5 0.04495 -0.03731
C5 C5 0.12206 0.01495
C4 C4 0.01664 -0.07078
C6 C6 -0.06819 0.08515
D2 D2 -0.12681 0.07048
D1 D1 -0.07156 0.09227
B4 B4 -0.04687 0.00885
N_A2 -0.03669 0.22140
N_A5 -0.08943 0.22104
N_A3 -0.08365 0.14676
N_A4 -0.09034 0.12082
N_A1 0.07652 0.06959
C3 C3 0.02205 -0.18024
C2 C2 -0.17152 -0.12543
C1 C1 0.16108 -0.13926

Proportion of variance explained by each factor

Obs Factor1 Factor2 k sumF w1 w2
1 5.7984077 4.9049874 1 10.7034 0.54174 0.45826

Means and variances of variables

Obs _TYPE_ B1 B2 B3 B4 B5
1 MEAN 98.5564 240.058 120.029 29.0867 0.18001
2 STD 44.0645 135.296 71.095 16.9387 0.07688

157
Obs C1 C2 C3 C4 C5 C6
1 0.15846 81.2374 46.5639 20.9006 8.08366 0.082161
2 0.09822 9.1263 23.4846 0.6053 2.44772 0.049863

Obs N_A1 N_A2 N_A3 N_A4 N_A5 D1 D2
1 1.52976 2.38097 1.93719 1.79475 1.78352 449.344 126.575
2 0.56578 0.87092 0.40464 0.48876 0.63802 87.989 95.598


158
Appendix 5. Singular Value Decomposition (SVD)
A description of the singular value decomposition (SVD) of a matrix can be found in
any basic text book on linear algebra (see for example Lay, 1994). In this appendix
only a very short overview will be given.

The SVD of a matrix A involves two orthogonal matrices and a matrix containing the
singular values of A. Therefore, the definitions of orthogonality and singular values
are given before the SVD is introduced.

Firstly, a matrix U is called orthogonal if it is a square matrix with orthonormal
columns. With other words, its columns form an orthogonal set of unit vectors. From
the definition it immediately follows that an nn orthogonal matrix U satisfies the
equality U
T
U = UU
T
= I
n
where I
n
is the nn identity matrix.

Secondly, the singular values of an mn matrix A are the square roots of the
eigenvalues of A
T
A. A scalar is by definition an eigenvalue of A
T
A if there exists a
non-zero n-dimensional vector x such that A
T
Ax = x. Each nn matrix has exactly n
eigenvalues (counting multiplicities), from which is follows tha each m n matrix has
exactly n singular values (also counting multiplicities).

Although the two definitions above are sufficient to define the SVD of an m n matrix
A, it is helpful to introduce some notation first. The singular values of will be
denoted by
1

2
...
r
> 0 =
r+1
= ... =
n
, so A has exactly r
singular values. The matrix D will be the
t

A
strictly positive
rr diagonal matrix with the strictly positive
singular values of A on the diagonal and will be the m n matrix given by


It follows that has m-r rows containing only zeros and n-r columns containing only
zeros.

The matrix plays an important role in the SVD of an m
.
0 0
0

=
D

n matrix A. Indeed, the
SVD of A is given by
(1)

where U is an m
.
T
V U A =
m orthogonal matrix and V an nn orthogonal matrix. There
always exist orthogonal matrices U and V such that equality (1) holds. Moreover,
they are not unique For each choice of U and V such that the previous equation
holds, the columns of U are called the left singular vectors of A, whereas V are
called the right singular vectors of A.

Because possibly contains zero columns and rows, the SVD can be reduced to
the following expression:


where U
r
and V
r
are the matrices consisting of the first r columns of U and V
respectively. Working out this expression shows that A can be written as the follow-
ing sum of matrices:
,
T
r r
DV U A =
159


Here u
i
and v
i
are the i-th columns of U
r
and V
r
respectively. The matrix A
i
is called
the i-th component of the SVD. Each component A
i
can be written out explicitly:

(2)

where u
ij
and v
ij
denote the j-th element of u
i
and v
i
respectively. It follows that the
rows of A
i
are multiples of the row vector v
i
T
and the columns of A
i
are multiples of
the vector u
i
.

The singular value
i
is a measure of the contribution of matrix A
i
to A. Because the
singular values are in decreasing order, A
1
contributes the most to A, i.e., it explains
most of the variance in the values of A. The cumulative percentage of explained
variance of the k-th component, denoted by CPEV
k
, is the variance of A explained
by It is computed as

. where ,
2 1
T
i i i i r
v u A A A A A = + + + = K
,
2 1
2 2 2 1 2
1 2 1 1 1

=
in im i im i im
in i i i i i
in i i i i i
i i
v u v u v u
v u v u v u
v u v u v u
A
L
M O M M
L
L

.
1 k
A A + +K
( ) ( )
( )
. 1
1 1
2
) , (
1 1
2
) , ( 1 ) , (

= =
= =
+ +
=
m
i
n
j
j i
m
i
n
j
j i k j i
k
A
A A A
CPEV
K


If the CPEV
k
is very large, then A is reasonably well approximated by

In this report the SVD is used to study the similarities and dissimilarities between the
development of the fatality rates over the years 1970 up to 2003 for the nine
countries. The matrix A represents in this case a table with in its (i,j)-entry the fatality
rate of country i in year j.

If A is reasonably well approximated by A
1
, then the development over time of the
fatality rate in the nine countries is similar. In other words, for each country the 34-
dimensional row vector containing the fatality rates over the years is equal to the
row vector
1
v
1
T
multiplied with a scalar. For country j this scalar is u
1j
. The row
vector v
1
T
is called the general trend of the SVD and the scalars u
1j
the country
weights of the first component. Analysing the country weights gives an idea of the
road safety in a country. Countries with larger weights perform worse than countries
with smaller weights.

The terminology introduced above can be extended if more components are added
to the approximation of A. The row vector v
j
T
is called the trend of the j-th compo-
nent and the scalar u
ij
is called the country weight of country i of the j-th component.
Adding more SVD components to the approximation of A makes it possible to study
in which way a country deviates from the general trend. If country i has the highest
country weight (in absolute value) of component j, then the trend of component j
primarily represents the deviations of the combined trend of the first j-1 components
for country i.
.
1 k
A A + +K
160
Appendix 6. The latent risk model
For each country a bivariate local linear trend model called the latent risk model was
used to estimate macroscopic trends and forecasts for the developments of road
safety and exposure. Whereas in a classical regression model the intercept and the
regression coefficient of the linear regression of a dependent variable on time are
fixed and do not change over time, in a local linear trend model these two
parameters are typically allowed to change from time point to time point. In this
context the time-varying intercept and regression weight are called the level and the
slope component, respectively.
The latent risk model is a special case of the state space methods for the analysis of
time series (Harvey, 1989; Durbin & Koopman, 2001; Commandeur & Koopman,
2007). In matrix algebra, all state methods can very generally be written as

t t t t
Z y + = , ) , 0 ( ~
t t
H NID (1)

t t t t t
R T + =
+1
, ) , 0 ( ~
t t
Q NID (2)

for t = 1, ..., n, where ) , 0 ( ~
t t
H NID is a short-hand notation for: the errors or
disturbances
t
are assumed to be normally and independently distributed with
means equal to zero and a variance structure equal to . The latent risk model for
evaluating the developments in road safety is a bivariate local linear trend model.
Specifically, let

,

where M
t
are the observed annual mobility figures and F
t
are the observed annual
fatality figures at time points t = 1, ..., n.
Defining

, , ,

, ,

t
H

=
t
t
t
t
t
F
M
y
y
y
log
log
) 2 (
) 1 (

=
) 2 (
) 2 (
) 1 (
) 1 (
t
t
t
t
t

=
) 2 (
) 2 (
) 1 (
) 1 (
t
t
t
t
t

=
1 0 0 0
1 1 0 0
0 0 1 0
0 0 1 1
T

=
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
R

=
0 1 0 1
0 0 0 1
Z
161

=
2
) 2 (
) 2 ( ) 1 (
) 2 ( ) 1 ( 2
) 1 (
) , cov(
) , cov(

t
H , and


and writing out (1) in scalar notation yields the following two observation equations:

, (3)

while working out (2) in scalar notation results in the following four state equations:

=
2 ) 2 ( ) 1 (
2 ) 2 ( ) 1 (
) 2 ( ) 1 ( 2
) 2 ( ) 1 ( 2
) 2 (
) 2 (
) 1 (
) 1 (
0 ) , cov( 0
0 0 ) , cov(
) , cov( 0 0
0 ) , cov( 0

t
Q
) 2 ( ) 2 ( ) 1 ( ) 2 (
) 1 ( ) 1 ( ) 1 (
t t t t
t t t
y
y


+ + =
+ =
) 2 ( ) 2 ( ) 2 (
1
) 2 ( ) 2 ( ) 2 ( ) 2 (
1
) 1 ( ) 1 ( ) 1 (
1
) 1 ( ) 1 ( ) 1 ( ) 1 (
1
t t
t
t t t
t
t t
t
t t t
t




+ =
+ + =
+ =
+ + =
+
+
+
+
. (4)

Since and are the trends for the exposure and the risk, respectively, and
the mobility and fatality figures are modelled in their logarithms, the second equation
in (3) can be written as


) 1 (
t

) 2 (
t

) error log( ) Risk trend log( ) Exposure trend log( log + + =

t
F (5)

and therefore as

, (6)

since
)] error )( Risk trend )( Exposure trend log[( log =
t
F
) log( log log ab b a = + .

Finally, taking the exponent of (6) yields the following multiplicative model

Traffic Safety = (trend Exposure)(trend Risk)(error) . (7)

When all disturbances in (4) are fixed on zero, the latent risk model collapses to a
bivariate linear regression model:

162
) 2 ( ) 2 ( ) 2 ( ) 1 ( ) 1 (
) 1 ( ) 1 ( ) 1 (
log
log
t t
t t
t b a t b a F
t b a M

+ + + + =
+ + =
(8)

with t = 1, , n.

As (7) clearly indicates, in the latent risk model the development of road safety is
assumed to be the product of the developments of two latent, unobserved factors:
exposure and risk, see also Bijleveld et al. (2008). The model requires the
estimation of thirteen parameters: three for the disturbance variances of the mobility
and fatality figures including their covariance, another three for the disturbance
variances of the level components of exposure and risk including their covariance,
yet another three for the disturbance variances of the slope components of exposure
and risk including their covariance, and finally four parameters for the initial values
of the two level and the two slope components.

In all cases the annual total number of road fatalities in each country was used as
an indicator for road safety. Whenever available the annual total number of motor
vehicle kilometres driven was used as an indicator for the exposure in a country; if
not available the annual figures for the total number of motor vehicles in a country
were used for this purpose instead.


163
Appendix 7. Fatality trends for individual countries
Observed and predicted fatality numbers for 20 European countries, using motor
vehicle kilometre figures (unless otherwise stated) and fatalities in Harvey's
structural time series model.

1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1500
2000
2500
3000
Fatalities: Austria
1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1500
2000
2500
3000
3500
Fatalities: Belgium
Austria Belgium
1980 1985 1990 1995 2000 2005 2010
600
800
1000
1200
1400
1600
Fatalities: Czech Republic
1970 1975 1980 1985 1990 1995 2000 2005 2010
200
400
600
800
1000
1200
1400
Fatalities: Denmark
Czech Republic: with vehicle fleet as indicator for exposure. Denmark: with mv-kms missing for 2003, 2005, and 2006.
1975 1980 1985 1990 1995 2000 2005 2010
300
400
500
600
700
800
900
Fatalities: Finland
1970 1975 1980 1985 1990 1995 2000 2005 2010
5000
7500
10000
12500
15000
17500
20000
Fatalities: France
Finland France
164

1980 1985 1990 1995 2000 2005 2010
4000
6000
8000
10000
12000
14000
16000
Fatalities: Germany
1970 1975 1980 1985 1990 1995 2000 2005 2010
3000
4000
5000
6000
7000
8000
Fatalities: Great Britain
Germany Great Britain
1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1250
1500
1750
2000
2250
2500
Fatalities: Greece
1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1250
1500
1750
2000
2250
2500
Fatalities: Hungary
Greece: with vehicle fleet as indicator for exposure (missing in
1992).
Hungary: with vehicle fleet as indicator for exposure.
1980 1985 1990 1995 2000 2005 2010
12.5
15.0
17.5
20.0
22.5
25.0
27.5
30.0
32.5
35.0
Fatalities: Iceland
1970 1975 1980 1985 1990 1995 2000 2005 2010
300
350
400
450
500
550
600
650
Fatalities: Ireland
Iceland: mv-kms missing for 2001, 2002, 2003, 2004, and
2005.
Ireland: with vehicle fleet as indicator for exposure.
1970 1975 1980 1985 1990 1995 2000 2005 2010
5000
6000
7000
8000
9000
10000
11000
12000
13000
14000
Fatalities: Italy
1970 1975 1980 1985 1990 1995 2000 2005 2010
40
60
80
100
120
Fatalities: Luxemburg
Italy: with vehicle fleet as indicator for exposure (and 2005
missing).
Luxemburg: with vehicle fleet as indicator for exposure.
165
166

1970 1975 1980 1985 1990 1995 2000 2005 2010
1000
1500
2000
2500
3000
3500
Fatalities: Netherlands
1975 1980 1985 1990 1995 2000 2005 2010
200
250
300
350
400
450
500
550
600 Fatalities: Norway
The Netherlands Norway
1990 1995 2000 2005 2010
4000
5000
6000
7000
8000
9000 Fatalities: Poland
1970 1975 1980 1985 1990 1995 2000 2005 2010
500
1000
1500
2000
2500
3000 Fatalities: Portugal
Poland: with vehicle fleet as indicator for exposure. Portugal: mv-kms missing for 2004, 2005, and 2006.
1970 1975 1980 1985 1990 1995 2000 2005 2010
200
300
400
500
600
700
800
Fatalities: Slovenia
1970 1975 1980 1985 1990 1995 2000 2005 2010
3000
4000
5000
6000
7000
8000
9000
Fatalities: Spain
Slovenia Spain: with vehicle fleet as indicator for exposure.
1970 1975 1980 1985 1990 1995 2000 2005 2010
400
600
800
1000
1200
1400
Fatalities: Sweden
1970 1975 1980 1985 1990 1995 2000 2005 2010
500
750
1000
1250
1500
1750
2000
Fatalities: Switzerland
Sweden Switzerland

167
Appendix 8. Disaggregate developments for
groups of countries
Development by age group
Group 1: 0-14 years
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
0,16
0,18
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Norway Sweden Average
Group 2: 0-14 years
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland
Italy Luxembourg Switzerland Average
Group 3: 0-14 years
0
0,01
0,02
0,03
0,04
0,05
0,06
0,07
0,08
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: 0-14 years
0
0,02
0,04
0,06
0,08
0,1
0,12
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

Group 1: 15-24 years
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Norway Sweden Average
Group 2: 15-24 years
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland
Italy Luxemburg Switzerland Average
Group 3: 15-24 years
0
0,05
0,1
0,15
0,2
0,25
0,3
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: 15-24 years
0
0,05
0,1
0,15
0,2
0,25
0,3
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

168

Group 1: 25-64 years
0
0,1
0,2
0,3
0,4
0,5
0,6
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Norway Sweden Average
Group 2: 25-64 years
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland
Italy Luxemburg Switzerland Average
Group 3: 25-64 years
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: 25-64 years
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

Group 1: 65+ years
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Norway Sweden Average
Group 2: 65+ years
0
0,05
0,1
0,15
0,2
0,25
0,3
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland
Italy Luxemburg Switzerland Average
Group 3: 65+ years
0
0,05
0,1
0,15
0,2
0,25
0,3
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: 65+ years
0
0,05
0,1
0,15
0,2
0,25
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

169
Developments by traffic mode
Group 1: motorized 2-wheelers
0
0,05
0,1
0,15
0,2
0,25
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Iceland
Netherlands Norway Sweden Average
Group 2: motorized 2-wheelers
0
0,05
0,1
0,15
0,2
0,25
0,3
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: motorized 2-wheelers
0
0,05
0,1
0,15
0,2
0,25
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: motorized 2-wheelers
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

Group 1: motorized 4-wheelers
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Iceland
Netherlands Norway Sweden Average
Group 2: motorized 4-wheelers
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: motorized 4-wheelers
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: motorized 4-wheelers
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average

170

Group 1: non-motorized
0
0,1
0,2
0,3
0,4
0,5
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Iceland
Netherlands Norway Sweden Average
Group 2: non-motorized
0
0,1
0,2
0,3
0,4
0,5
0,6
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: non-motorized
0
0,1
0,2
0,3
0,4
0,5
0,6
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: non-motorized
0
0,1
0,2
0,3
0,4
0,5
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Greece Portugal Spain Average
Developments by location
Group 1: country roads
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Sweden Average
Group 2: country roads
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: country roads
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: country roads
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Portugal Spain Average

171

Group 1: urban roads
0
0,1
0,2
0,3
0,4
0,5
0,6
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Sweden Average
Group 2: urban roads
0
0,1
0,2
0,3
0,4
0,5
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: urban roads
0
0,1
0,2
0,3
0,4
0,5
0,6
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: urban roads
0
0,1
0,2
0,3
0,4
0,5
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Portugal Spain Average

Group 1: motorways
0
0,02
0,04
0,06
0,08
0,1
0,12
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Denmark Finland Great Britain Netherlands Sweden Average
Group 2: motorways
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
0,16
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Austria Belgium France Germany Ireland Italy Switzerland Average
Group 3: motorways
0
0,02
0,04
0,06
0,08
0,1
0,12
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Czech Republic Hungary Poland Slovenia Average
Group 4: motorways
0
0,02
0,04
0,06
0,08
0,1
1970-1974 1975-1979 1980-1984 1985-1989 1990-1994 1995-1999 2000-2004 2005-2006
P
r
o
p
o
r
t
i
o
n

o
f

r
o
a
d

f
a
t
a
l
i
t
i
e
s
Portugal Spain Average

172
Appendix 9. IAL - Local Accident Indicator
The Local Accident Indicator (Indicateur dAccidentologie Locale: IAL) is computed
as the ratio of the number of deaths observed in the dpartement under study to the
number of deaths that would have been recorded if the risk exposures, by road
category, had been identical to those measured in France nationwide.

Five road categories are defined: (1) urban units of 5,000 + inhabitants; and outside
these urban units: (2) highways (autoroutes), (3) main roads (routes nationales), (4)
local roads (routes dpartementales), (5) 'other roadways', typically urban roads.

The risk exposure in category (1) is defined as the ratio of persons killed in an
accident occurring in an urban unit of 5,000+ inhabitants to the total population of
these units, expressed in tens of thousands of inhabitants (on the basis of the latest
population census). For categories (2), (3), and (4), we calculate the ratio of the
number of fatalities to the total mileage travelled in each of these networks,
expressed in hundreds of millions of kilometres. For the last category (5), the
population is used as a proxy for the traffic.

These risk levels are computed identically for all geographic areas: dpartement,
region, and all of France. The number of fatalities used is the total of the past five
years, recalculated annually on a sliding basis. The choice of a structural indicator
was guided by the desired objective: what we want to measure is not so much a
short-term pattern as the stable notion of relative risk. For the sake of consistency,
the mileage estimates are based on data (network length and mean daily flow) for
the same five-year period. Therefore, the estimated values are annual averages.

To understand this properly, let us take the (fictitious) example of a dpartement in
which 500 people have been killed in the past five years, of whom 90 in an urban
unit of 5,000+ inhabitants, 25 on highways, 110 on main roads, 250 on local roads,
and 25 on 'other roadways.' The nationwide risk levels for these five roadway
categories are respectively 1.76, 0.53, 1.99, 2.04, and 2.26. For the dpartement
observed, they stand at 3.39, 0.48, 2.66, 2.51, and 2.45. If the risk exposure in the
dpartement under study had been identical to the total French levels for each
roadway category, the number of fatalities recorded in the dpartement would have
been:

90*(1.76/3.39) + 25*(0.53/0.48) + 110*(1.99/2.66) + 250*(2.04/2.51) + 25*(2.26/2.45)
or 47 + 28 + 82 + 203 + 23 = 383.

The 'overall' local accident indicator for our control dpartement is equal to 500/383,
or 1.31. For highways alone, it is equal to 25/28 = 0.89.


173
Appendix 10. Literature overview of studies on
regional risk analysis
Author, yearP
1
P Country Level Measure Method Type Dimension
Hindle et al.,
2008
England Regions Counts,
KSIP

P
Regression
model
Descriptive Time
Eksler et al.,
2008a, b, c
EU-25,
Belgium,
Czech
Republic,
France
Regions,
Municipalities
Mortality,
Fatality
rates,
Accident
rates
Full Bayes Descriptive,
Ecological
Space,
time
La Torre et
al., 2007
Italy Regions Mortality
rates
Basic
rates,
Multiple
regression
Ecological Space
Aguero-
Valverde &
Jovanis, 2006
Pennsylvania Communes Fatal and
injury
accident
rates
Full Bayes Ecological Space-
time
Broughton &
Buckle, 2006
England NUTS-3 Fatalities,
KSI
Basic Descriptive Time
Lassarre &
Thomas,
2005
17 EU
countries
NUTS-2 Mortality
rates
Empirical
Bayes
Descriptive Space
Haynes et al.,
2005
England and
Wales
Districts Fatality
counts
Odds ratio,
Basic rates
Descriptive Space
Noland &
Quddus,
2004
Illinois and
England
Regions Fatality
and injury
counts
Full Bayes
MacNab,
2004
British
Columbia
Regions Mortality
rates
Full Bayes Ecological Space
Amoros et al.,
2003
France Counties Mortality
and fatality
rates
Basic Descriptive Space
Shaw et al.,
2000
EU-15 NUTS-2 SMRP
2
P Basic Descriptive Space
Williams et
al., 1991
Scotland NUTS-3 SMRP
2
P Basic Descriptive Space
Van Beeck,
1991
Netherlands Regions Fatality,
injury rates
Basic Ecological Space
Fridstrom &
Ingebrigtsen,
1991
Sweden,
Norway,
Finland
Counties Accident,
fatality
counts
GLMP
3
PPP Ecological Space
Baker et al.,
1987
New England Regions Mortality
rates
Basic
rates,
Regression
analysis
Ecological Space-
time
1) For the full references: see next page.
2) SMR = Standardized mortality ratio.
3) GLM = Generalized linear model.

174
Full references to Appendix 10

Amoros, E., Martin J.L. & Laumon, B. (2003). Comparison of road crashes incidence and severity
between some French counties. In: Accident Analysis and Prevention, vol. 35, nr. 4. p. 537547.

Aguero-Valverde, J. & Jovanis, P.P. (2006). Spatial analysis of fatal and injury crashes in
Pennsylvania. In: Accident Analysis and Prevention, vol. T38, nr. T3, p. 618-625.

Baker, S., Whitfield, R.A. & ONeill, B. (1987). Geographic variations in mortality from motor vehicle
crashes. In: The New England Journal of Medicine, vol. 316, nr. 21, p. 1384-1387.

Broughton, J. & Buckle, G. (2006). Monitoring progress toward the 2010 casualty reduction target -
2004 data. TRL Report 653. Transport Research Laboratory TRL, Crowthorne, Berkshire.

Eksler, V., Lassarre, S. & Thomas, I. (2008a). The regional analysis of road mortality in Europe: A
Bayesian ecological regression model. In: Public Health, vol. 122, nr. 9, p. 826-837.
doi:10.1016/j.puhe.2007.10.003

Eksler, V. , Lassarre, S. & Thomas, I. (2008b). Belgian regions: Federated or divided in road fatality
risk evolution? In: 48th Congress of the European Regional Science Association ERSA, 27-31 August
2008, Liverpool.

Eksler, V. & Lassarre, S. (2008c). Evolution of road risk disparities at small-scale level: example of
Belgium. In: Journal of Safety Research, vol. 39, nr. 4, p. 417-427.

Fridstrom, L. & Ingebrigtsen, S. (1991). An aggregate accident model based on pooled, regional time-
series data. In: Accident Analysis and Prevention, vol. 23, nr. 5, p. 363-378.

Haynes, R., Jones, A., Harvey, I., Jewel, T. & Lea, D. (2005). Geographical distribution of road traffic
deaths in England and Wales; Place of accident compared with place of residence. In: Journal of
Public Health, vol. 27, nr. 1, p. 107-111. doi:10.1093/pubmed/fdh212

Hindle, G.A., Hindle, T. & Souli S. (2008). Modelling and assessing local area differences in road
casualties: a case study in England. In: Journal of the Operational Research Society, in press.

La Torre, G., Van Beeck, E., Quaranta, G., Mannocci, A. & Ricciardi, W. (2007). Determinants of
within-country variation in traffic accident mortality in Italy: a geographical analysis. In: TInternational
Journal of Health GeographicsT, Tvol. 6: T49 (8 pp). Tdoi:10.1186/1476-072X-6-49T

Lassarre S. & Thomas, I. (2005). Exploring road mortality ratios in Europe: national versus regional
realities. In: Journal of the Royal Statistical Society Series A, vol. 168, part 1, p. 127-144.

MacNab, Y.C. (2004). Bayesian spatial and ecological models for small-area accident and injury
analysis. In: Accident Analysis and Prevention, vol. T36, nr.T 6, p. 10281091.

Noland, R.B. & Quddus, M.A. (2004). A spatially disaggregate analysis of road casualties in England.
In: Accident Analysis and Prevention, vol. 36, nr. 6, p. 973-984.

Shaw, M., Orford, S., Brimblecombe, N. & Dorling, D. (2000). Widening inequality in mortality between
160 regions of 15 European countries in the early 1990s. In: Social Science and Medicine, vol. T50T, nrs.
7-8, p. 1047-1058.

Van Beeck, E.F., Mackenbach, J.P., Caspar, W.N., Looman, N. & Anton, E. (1991). Determinants of
traffic accident mortality in the Netherlands: a geographical analysis. In: International Journal of
Epidemiology, vol.20, nr. 3, p. 698-706.

Williams, F.L.R., Lloyd, O.L. & Dunbar, J.A. (1991). Deaths from road traffic accidents in Scotland:
(1979-1988). Does it matter where you live? In: Public Health, vol. 105, nr. 4, p. 319-326.