category: literaturenote citekey: gitelmandesigningcompositeindicatorroad2010 title: Designing a composite indicator for road safety authors: "Gitelman, Victoria; Doveh, Etti; Hakkert, Shalom" year: 2010 date: 2010-00-00 2010 doi: 10.1016/j.ssci.2010.01.011 publication: Safety Science url: "https://doi.org/10.1016/j.ssci.2010.01.011" zotero_key: B4W7HUXZ zotero_storage: 9DKWFUXN collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Gitelman, Victoria"

status: converted

Contents lists available at ScienceDirect

Safety Science

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

Designing a composite indicator for road safety

Victoria Gitelman a,*, Etti Doveh ^b , Shalom Hakkert ^a

^a Transportation Research Institute, Technion City, Haifa 32000, Israel
^b The Technion Statistical Laboratory, Technion City, Haifa 32000, Israel

article info

Article history: Received 1 September 2009 Received in revised form 12 January 2010 Accepted 14 January 2010

Keywords: Road safety Performance Composite indicator Statistical weighting

abstract

This study set out to develop a composite road safety indicator for benchmarking countries' road safety performance, which would combine the main layers of the road safety pyramid which describes the complex nature of road safety activities, performance and outcomes. Four groups of basic safety indicators were considered, which refer to: policy performance (road safety programmes), final road safety outcomes (fatality rates, scope of traffic injury), intermediate outcomes (wearing rates of seat belts, crashworthiness and composition of vehicle fleet, alcohol-impaired driving), and background characteristics of countries (motorization level, population density). The analysis used the data collected for 27 European countries. Weights based on statistical models were used to combine the basic indicators into a composite one. Principal Component Analysis and Common Factor Analysis weighting were examined. The composite indicators, estimated by several methods, enabled us to rank and group the countries according to their safety performance.

The analysis revealed that the countries' ranking based on the composite indicators is not necessarily similar to the traditional ranking of countries based on fatality rates only. Furthermore, it was observed that the indicators belonging to the final outcomes and intermediate outcomes are not uniform in their behaviour. Indicators which were found to be more consistent and influential and termed a 'core set of basic indicators' are recommended for future uses. The general conclusion is that the design of a composite road safety indicator in which relevant information from the different components of the road safety pyramid has been captured and weighted is realistic and meaningful. Such an indicator gives a more enriched picture of road safety than a ranking based only on fatality rates, which is the common practice at present. Grouping countries in this process is promising and seems to be preferable to simply ranking countries.

-2010 Elsevier Ltd. All rights reserved.

1. Introduction

The number of indicators suggested for use in the field of road safety has been growing rapidly, especially over the last decade (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 over 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 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 indicator which can describe all the relevant components in a concise and comprehensive way. 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 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 (Al Haji, 2005; Nardo et al., 2005).

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 started with the definition of eight

^* Corresponding author. Tel.: +972 4 8295132; fax: +972 4 8225716. E-mail addresses: trivica@tx.technion.ac.il (V. Gitelman), ierde01@ie.technion. ac.il (E. Doveh), hakkert@tx.technion.ac.il (S. Hakkert).

dimensions of the road safety domain, which are traffic risk, personal risk, vehicle safety, road situation, road user behaviour, socio-economic background, road safety organisation and enforcement. For each dimension, one or several quantitative indicators were suggested and their applicability was analysed based on available data. Finally, each country was characterized by eleven separate indicators which were combined into one composite indicator. To construct a composite index, Al Haji (2005) applied three ways of weighting, which were the simple equal average, the use of theoretical weights, and 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 WHO, 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 risk (the number of traffic fatalities per million inhabitants). It was found that different weighting 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 risk.

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.

In this study, we aimed to create and explore a comprehensive composite road safety indicator, 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. In the SafetyNet project safety performance indicators were developed, but no attempt was made to combine the results from different indicators into one composite value, in order to compare countries. Somewhat earlier the SUNflower project tried to develop a methodology to compare countries on their safety level. This resulted in the footprint methodology. It was felt that, on the basis of these prior methods and the available data, a new composite indicator could be developed, what has been carried out in the present study.

In this process, the types of benchmarking discussed in the literature were considered, and data for the European countries were used that were collected in the SafetyNet project and which were available from international databases. This study was a component of a comprehensive project aimed at developing a knowledge-based framework for the benchmarking of road safety performances and developments for national and sub-national jurisdictions (Wegman et al., 2008).

2. The model

2.1. The road safety pyramid

The SUNflower approach (Koornstra et al., 2002; Wegman et al., 2005) 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 layer of safety performance indicators (SPIs) was originally explored by ETSC (2001). The reason for the SPIs development was 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. Thus, SPIs can be defined 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. Based on the potential of different road safety areas 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. They are: (1) alcohol and drug-use; (2) speeds; (3) protective systems; (4) daytime running lights; (5) vehicles (passive safety); (6) roads and (7) trauma management.

For each one of these areas, SPIs were developed and, then, using the data provided by national representatives of the European countries, SPIs were estimated and country comparisons were performed – see Vis and Van Gent (2007).

2.2. Basic indicators

This study set out to develop a composite road safety indicator for benchmarking purposes, which would combine the main layers of the road safety pyramid.

The layers considered were the road safety policy performance (safety measures and programmes), accident fatalities/injuries (as the final outcomes) and safety performance indicators (as intermediate outcomes). These layers will be further described by A-, Band C-groups of indicators, respectively. 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').

To develop a composite indicator, basic indicators should first be defined, for each layer of the pyramid that needs to be considered. As stated earlier, the present study builds upon the theory and evidence developed and collected in two earlier related projects: the SUNflower (Wegman et al., 2005) and SafetyNet (Hakkert et al., 2007) studies. In these studies a theory was developed linking the characteristics of four basic layers related to safety, especially concerning the final outcomes' and intermediate outcomes' layers. In the SafetyNet project a large amount of evidence was collected and described linking the indicators of intermediate outcomes to safety. On the other hand, the final selection of basic indicators was limited by the data currently available for the European countries from international databases and from the Safety-Net project.

As for the policy indicators, results have been reported in the literature that countries with comprehensive safety programs and quantitative targets usually performed better than others. The main principles of successful road safety management are well summarized by OECD (2002), ETSC (2006), Elvik (2008). Based on these principles, for the A-group of indicators a number of characteristics were suggested which would reflect the quality of national road safety programmes. The five indicators introduced for the Agroup related to the level of national safety targets, the approach applied for selection of interventions, the scope of economic evaluation which preceded the programme's composition, the way of monitoring and the level of responsibility for the programme's performance. For each indicator, the categories of values were defined as presented in Table 1.

For each country, the value of each indicator should be selected based on the published national road safety programmes, other background papers and available follow-up reports, including those prepared by the European Transport Safety Council (ETSC), Organisation for Economic Co-operation and Development (OECD) or other international working groups. The values applied in the current study should be treated as preliminary as those are mostly based on two sources: country reports from OECD/ECMT (2008) and surveys provided by OECD/ITF (2008), and on some other frequently cited road safety publications.

Giving the values according to Table 1, the national safety programme was marked as having an ''ambitious" target when it was in line with the European policy of 50% reduction in fatalities by 2010 or when it implied an average annual reduction of fatalities by 5% or more; such a mark (value ''a") was given for many countries. On the other hand, some countries (e.g. Italy) stated that they have no national targets (value ''c"); the remainder of countries received value ''b".

Considering the selection of safety interventions for the national programme, the best mark (value ''a") was given for the programmes where a comprehensive categorization of road safety problems and estimation of their magnitude was performed as it was done, for example, for Sweden by Elvik and Amundsen (2000) and for Switzerland by VESIPO (2002). Value ''b" of this indicator reflects the emphasis on evidence-based measures selected by the programme, where the ''c" value indicates a case where detailed analysis of accident statistics is performed but no estimation of safety effects associated with the measures is provided.

Considering the economic evaluation, the best mark (value ''a") is given for the programme, when a background study estimated the safety benefits expected. For example, in Switzerland, the development of a road traffic safety policy was based on a comprehensive study which assessed the efficiency of 77 road safety measures (VESIPO, 2002); Elvik (2001) demonstrated cost-effective analyses of road safety policies for Norway and Sweden, which enveloped 132 and 139 road safety measures, respectively. Usually, a general economic evaluation is performed where the numbers of fatalities and injuries to be saved are compared with the investments required for the programme's performance (value ''b").

For the monitoring indicator, value ''a" is assigned when the monitoring analysis reports are available, e.g. Broughton and Buckle (2006) for the UK. As to the stakeholders' indicator, the level of adoption and responsibility for the programme's performance can be learnt from the OECD country reports where the governmental responsibility (value ''a"), the responsibility of national road safety authority (value ''b"), shared responsibility of several bodies (value ''c") or no clear responsibility (value ''d") are indicated.

In spite of the thorough effort to provide the data, it should be realised that it is difficult to obtain quantitative and unbiased estimates on the actual implementation of road safety policies and programs.

The B-group of indicators deals with the final outcomes of the system, i.e. the numbers of road accident fatalities and injuries, which should be presented in a form suitable for comparisons. The present study focused on four issues: personal risk, traffic risk, the scope of traffic injury and the scope of the problem of vulnerable road users. For each issue one or several indicators were defined – Table 2.

Concerning this group of indicators, we can note that personal risk and traffic risk indicators are common in international comparisons; the indicators of traffic injury, in addition to fatalities, receive more attention over the last years both in international comparisons and in setting national safety targets (e.g. OECD/ITF, 2008). As to the share of pedestrians/two-wheelers out of the total fatalities, those are compared among the countries where higher than average values frequently lead to more concerted safety efforts and interventions related to the subject (e.g. pedestrian safety issue in the national safety programme in the UK).

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 available in the seven pre-defined SPI areas (Vis and 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 EU Member States. The remaining safety areas are: alcohol-impaired driving, use of protective systems in cars, the crashworthiness of the passenger car fleet, and the vehicle fleet composition. For each area one or two indicators have been defined – Table 3.

Table 1 Definition of basic indicators for the A-group: characteristics of national safety programmes.

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. Detailed economic evaluation preceded the programme's composition
	b. Some economic evaluation was performed
	c. Economic evaluation was not performed
A4 Monitoring the programme's	a. Systematic monitoring takes place
performance	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 2 Definition of basic indicators for the B-group: final outcomes.

Issues considered	Indicators defineda
Personal risk	B1 Fatalities per million inhabitants
Traffic risk	B2 Fatalities per million passenger cars
	B3 Fatalities per 10 billion passenger-km travelled
Scope of traffic injury	B4 Injury accidents per fatalityb
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

^a Note: Obviously, various definitions are possible. B2 can for instance be defined per total vehicle fleet instead of passenger cars, B3 per vehicle kilometer travelled instead of passenger kilometer, etc. The selection was based on the estimates available to the EU Member States.

Table 3 Definition of basic indicators for the C-group: intermediate outcomes, SPIs.

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 seat belts in the front seats (aggregated for driver and front passenger)
	C3 Daytime wearing rates of seat belts in the rear seats
Vehicles: crashworthiness of the passenger car fleet	C4 Average EuroNCAP score of passenger car fleet
	C5 Median age of the passenger car fleet
Vehicle fleet composition	C6 Share of motorcycles in the vehicle fleet
	C7 Share of heavy goods vehicles (HGV) in the vehicle fleet

Finally, two background indicators were added (the D-group) which characterize the motorization level and the population density of the country – Table 4. Both indicators are common in international and other comparisons as context variables reflecting the development and structural differences among the countries or other entities compared (Wegman et al., 2008).

The choice of indicators used was influenced by the immediate availability of data and, therefore, their definitions have a preliminary character, supporting the main purpose of the study, to demonstrate along which lines composite indicators can be designed.

2.3. Data collection

In total, 21 basic safety 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 and Van Gent (2007); for C5 – UNECE (2008); for C6, C7 – EC (2007) and OECD country reports (2008).

The basic indicators were estimated for 27 countries: 25 Member States of the European Union plus Norway and Switzerland – Table 5. (The majority of indicators are not yet available for the new Member States Bulgaria and Romania.)

2.4. Method of analysis

The purpose of this analysis is to create a composite road safety indicator and to explore the behaviour of basic indicators. The

Table 4 Definition of basic indicators for the D-group: background characteristics.

Characteristic considered	Indicators defined
Motorization level	D1 Number of passenger cars per 1000 inhabitants
Population density	D2 Population per 1 km2 of country's territory

composite indicator 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 5, only six countries had values for the whole set of basic indicators A–D. Therefore, prior to the analysis, missing data imputations were performed. The SAS 9.2 MI procedure was applied using the MCMC method for imputation. This imputation assumes multivariate normality of variables. We used SAS macro % bctrans to find a Box–Cox transformation in order to achieve marginal normality of each of the continuous variables (LaLonde, 2005). Due to the relatively small number of observations with respect to the number of basic indicators, the imputations were carried out separately for the groups of indicators. The final data set, with initial and imputed values of basic indicators, is presented in Appendix A.

To combine the basic indicators into a composite one, weights based on statistical models were applied. We examined both Principal Component Analysis (PCA) and Common Factor Analysis (FA) weighting (Tabachnick and Fidell, 2006). Both methods group collinear indices to form a composite indicator 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. This implies that in both methods the number of factors fitted to the data were data-dependent and not dictated by an external reason.

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.

^b Considering the reporting problems and definition's differences among the countries, a preferable indicator would be the 'number of hospitalized injuries per one fatality', which is unavailable as yet.

Table 5Basic indicators estimated for 27 European countries (n/a ⁼ not available).

Cou ntry	Cod	e B 1	B2	B3	B4	B5	B6	B7	D1	D2	C1	C2	C3	C4	C5	C6	C7	A1	A2	A3	A4	A5
		Fata litie s mil lion per inh abit ant s 200 6	Fata litie s per mil lion pas sen ger s-2 006 car	Fata litie s per 10 b illio n pkm 200 6	Inju ry Acc ide nts per fata litie s 200 6	Ped ian estr fata litie s as % of l tota 6 (% ) 200	Bicy clis t fata litie s as % of l tota 6 (% ) 200	Mo ycli torc st + ped mo fata litie s of t l as % ota 6 (% ) 200	Num ber of pas sen ger cars per 100 0 inh abit ant s, 200 6	Pop ulat ion m2 1 k per of a rea 200 6	Sha f re o drin k driv ing iden ts acc 200 5	We arin g -Fro rate nt ts sea d ate agg reg 200 5	We arin g rate Rea r ts sea 200 5	Ave rag e Eur oNC AP sco re 200 3	Me dia n of age pas sen ger s-2 003 car	% Mo ycle torc s in fl eet 6 (% ) 200	% HG V in flee t 200 6 (%)	Safe ty targ ets	Sele ctio f n o inte ntio rve ns	Eco ic nom luat ion eva	Mo nito ring	Sta ke hol der s
Aus tria	AT	88	175	100	54. 6	15. 1	6.6	18. 4	507	98. 9	6.0 %	83	52	n/a	7.5	12. 4	7.0	a	c	b	b	a
Belg ium	BE	101	216	96	38. 5	11. 4	8.6	15. 5	470	347 .0	4.4 %	71	n/a	21. 3	6.7	6.0	11. 1	b	b	b	b	b
lan Swi tzer	d C H	50	95	n/a	58. 1	20. 5	9.5	21. 6	519	181 .8	19. 3%	82	53	n/a	7.3	12. 5	6.4	a	a	a	b	a
Cyp rus	CY	111	236	167	31. 1	22. 1	2.3	29. 1	479	83. 8	22. 5%	80	n/a	n/a	9.3	7.6	21. 7	n/a	n/a	n/a	n/a	n/a
Cze ch	CZ	104	264	146	20. 8	19. 0	10. 3	10. 9	399	130 .4	5.5 %	72	13	20. 7	11	15. 1	9.0	b	c	b	b	b
Rep ubl ic
Ger ma ny	DE	62	110	57	64. 4	14. 0	9.5	17. 7	566	230 .6	5.1 %	96	89	20. 1	6	10. 5	5.1	a	b	b	a	b
rk Den ma	DK	56	154	56	17. 7	19. 6	10. 1	14. 7	371	126 .4	23. 0%	85	63	n/a	7.3	6.7	18. 7	a	b	b	b	b
nia Esto	EE	152	389	175	12. 7	27. 2	4.1	4.1	413	29. 7	28. 4%	74	30	n/a	12. 7	1.9	14. 0	b	c	b	b	a
Gre ece	EL	149	375	174	9.8	16. 1	1.3	30. 0	407	84. 6	10. 7%	n/a	n/a	20. 7	8	17. 2	17. 4	a	b	b	b	b
Spa in	ES	93	201	117	24. 3	15. 3	1.8	17. 6	464	87. 9	n/a	74	51	n/a	7.6	14. 6	16. 7	a	c	b	b	b
Fin land	FI	64	136	53	20. 1	11. 9	11. 3	9.5	475	15. 6	23. 5%	88	78	n/a	9.1	10. 5	11. 6	a	b	b	b	b
Fra nce	FR	77	153	64	17. 1	11. 4	3.8	23. 5	504	113 .1	28. 8%	97	70	n/a	8.3	6.4	13. 7	a	a	b	a	a
Hun gar y	HU	129	446	271	16. 1	22. 7	11. 7	10. 1	293	108 .2	8.8 %	67	34	20. 6	8	3.7	12. 5	a	c	b	b	b
Irel and	IE	86	211	130	18. 6	19. 0	3.0	16. 3	418	61. 4	n/a	86	46	n/a	5.1	1.6	14. 7	b	b	b	a	b
Ital y	IT	96	162	74	42. 0	12. 6	4.8	23. 8	597	196 .3	n/a	71	n/a	n/a	7.9	20. 5	8.7	c	d	c	c	d
Lith ia uan	LT	224	498	191	8.9	34. 0	n/a	n/a	470	51. 9	11. 8%	60	n/a	n/a	14. 5	1.4	7.7	b	c	b	b	b
Lux bou em rg	LU	76	116	54	21. 2	9.7	1.6	0.0	661	183 .1	0.0 %	80	60	n/a	4.4	10. 0	8.0	n/a	n/a	n/a	n/a	n/a
Latv ia	LV	178	520	256	10. 6	38. 0	6.0	4.0	360	35. 3	21. 7%	77	n/a	21. 3	10	3.7	12. 2	b	c	b	b	b
Ma lta	MT	25	46	49	84. 1	35. 3	0.0	17. 6	535	136 0.0	n/a	96	28	18. 6	9	4.4	16. 4	a	c	b	b	b
her lan ds Net	NL	45	102	48	33. 6	9.4	18. 3	18. 4	442	394 .2	15. 3%	90	64	20. 2	6	6.5	11. 3	a	a	b	a	a
Nor way	NO	52	116	n/a	32. 7	14. 5	3.3	15. 3	445	14. 4	n/a	91	83	21	8	9.4	17. 0	b	a	a	b	b
Pol and	PL	137	408	235	8.9	32. 3	11. 1	3.9	351	121 .9	9.8 %	78	n/a	n/a	11. 8	4.7	14. 4	b	c	b	b	b
al Por tug	PT	92	228	131	36. 8	16. 0	4.1	24. 1	405	115 .3	n/a	86	45	n/a	9.7	9.0	21. 3	b	c	b	b	b
Swe den	SE	49	107	45	40. 7	12. 4	5.8	15. 7	461	20. 3	34. 0%	92	73	22. 6	7	9.6	9.2	a	a	a	a	a
Slov eni a	SI	131	270	113	44. 4	13. 7	5.3	20. 5	488	99. 0	32. 2%	87	30	n/a	5.0	4.8	6.3	b	c	b	b	b
Slov akia	SK	107	439	215	13. 8	9.3	5.5	7.4	247	110 .5	12. 0%	n/a	n/a	n/a	n/a	3.7	11. 9	b	c	b	b	b
Gre at Brit ain	UK	54	116	48	58. 9	21. 0	4.6	17. 5	471	249 .3	5% 17.	90	84	20. 6	5	3.7	10. 7	a	b	b	a	a

In both methods (PCA/FA), the combination of factors into a composite indicator 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 of factors and, hence, 'the variance explained by each factor' was applied for estimating the weights.

Both PCA and FA are suitable for the creation of a composite indicator. Accounting for the data behaviour, five trials of creating a composite indicator were performed. They are:

(1) PCA-all A PCA in which all 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 indicator.
(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.

Concerning the FA analyses, 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), C7 (0.136). Once the indicators B6, B7, C7 were excluded, the overall KMO statistics increased to 0.63, enabling to proceed with the FA. Therefore, the three trials of creating composite indicator by means of FA considered 18 variables (basic indicators), without B6 (share of bicyclist fatalities), B7 (share of motorcyclist fatalities) and C7 (percentage of HGV in fleet).

To determine the number of factors required, the variance explained criterion was applied. At this point, two approaches are possible. Some researchers use the rule of keeping enough factors to account for a 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, for example, that four factors explain more than 80% of the variation.

Further, examining the solution for four factors demonstrated that the C4 variable (average EuroNCAP score) is nearly identical to one of the factors, where in such cases 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 the total explained variance than in the four factors' analysis (71% versus 84%). A further consideration revealed that keeping C4 inside the set and limiting the number of factors to two yielded a comparable value of the explained variance (about 69%). Hence, we decided to explore the results of all three trials as mentioned above.

Each of the five analyses produced a rotated factor pattern showing which variables (basic indicators) are more heavily loaded by each one of the factors built, i.e. the correlations between the variables and the factors. Considering the 'safety-desirable' behaviour of major variables which composed the factors and their coefficients, it was stated whether higher or lower values of each factor are associated with better safety performance. An insight into similarities and dissimilarities in the behaviour of basic indicators was also provided. In addition, each trial produced the tools for the estimation of each country's score, including: scoring coefficients, factors' weights, means and variances of the variables to estimate the standardized values. Appendix B provides an example of a rotated factor pattern and of tools for the estimation of country scores produced by the 'FA-4factors' analysis.

Once the factors are created, the following procedure is applied to calculate a composite indicator, 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 Factors 2, 3, etc. (if relevant). This creates the data set scores.
e. The composite indicator is a weighted sum of the Factors (called WF – weighted factor).

Using the WF and a WARD clustering procedure, in each trial a classification tree was produced, i.e. the countries were classified into similar groups (clusters). Inside the group the WF values are close, but there are distances between the groups.

3. Results and discussion

Detailed results of each of the five analyses can be found in Wegman et al. (2008). Section 3.1 gives a summary of the results of each trial. Furthermore, in Section 3.2, the composite indicators and clusters of countries are compared among the trials, providing considerations of their meanings and practical implications for the development of composite safety indicator.

3.1. Results of separate trials

3.1.1. PCA-all

In the 'PCA-all' analysis, five factors were fitted to the data (explaining 76.3% of the total variation), where

Factor 1 – Mainly reflects the road safety outcomes (the Bgroup), car fleet's age and seatbelt use;

Factor 2 – The policy performance indicators (the A-group) but also includes a negative correlation with C1 (share of drinkdriving accidents);

Factor 3 – Reflects the share of bicyclist fatalities, EuroNCAP scores for cars and population density;

Factor 4 – The share of motorcycles in the fleet and the share of motorcyclist fatalities;

Factor 5 – The share of HGV in the fleet, the number of injury accidents per fatality and the motorization level of a country.

Similar behaviour was noted for variable groups such as: B1-B2- B3 (number of fatalities per population, vehicles, passenger kilometers 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).

Fig. 1 presents the countries' subdivision into six clusters based on the WF values, where the country positions are plotted using the combined indicator (WF) and Factor 1 values. In this model, better safety position is associated with lower WF 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, and so on.

3.1.2. 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 combined safety index is generated.

For the A-group indicators ('policy performance'), one factor was fitted to all basic (A1–A5) indicators in the first step of this analysis (explaining 64.4% of the total variation).

For the B-group indicators ('final outcomes'), two factors were chosen, explaining 74.5% of the total variation, where variables B1-B2-B3-B4-B5 (number of fatalities per population, vehicles, passenger kilometers travelled, injury accidents per fatality, and share of pedestrian fatalities) were mainly loaded by Factor 1 and B6–B7 (shares of bicyclist and motorcyclist fatalities) – by Factor 2. The results demonstrated similarities in the behaviour of B1- B2-B3-B5 indicators and very different behaviours of B4, B6, B7.

Exploring the C-group indicators ('intermediate outcomes'), low 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, explaining 80.1% of the total variation, where variables C2–C3 (seat belt wearing rates) and C5 (median age of passenger cars) were mainly loaded by Factor 1, while C1 (share of drink-driving accidents) and C4 (average EuroNCAP score) were mainly loaded by Factor 2.

For the D-group indicators ('background characteristics'), one factor was chosen, which reflected both basic variables (explaining 63.2% of the total variation).

In the second step of this analysis two factors were chosen (explaining 86.3% of the total variation), where the combined factors of A- and C-groups were mainly loaded by Factor 1, and the combined factors of B- and D-groups – by Factor 2. Based on the factors' scoring coefficients and components, in this analysis, lower values of A- and B-group factors and higher values of C-group factors are 'safety-preferable', leading to lower values of Factor 1 and higher values of Factor 2 as associated with a better safety position.

Plotting the positions of countries on the dimensions of these two combined factors, one general 'cloud' can be noted, with several outliers (e.g. Italy, Malta) – Fig. 2. Inside the 'cloud', the countries with better safety-related positions are: France, Great Britain, Sweden, Switzerland, the Netherlands, Germany and Norway. Besides, in this analysis a lack of similarity in the behaviour of the group factors was recognized.

Fig. 1. Countries plotted using the composite indicator (WF) and Factor 1 values (PCA-all analysis).

3.1.3. FA with four factors

In the 'FA-4factors' analysis (see a rotated factor pattern in Appendix B) the basic indicators B1–B3 (number of fatalities per population, vehicles, passenger kilometers 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 population) were mainly loaded by 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 mainly loaded 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 programmes) and C1 (share of drink-driving accidents). The four factors explained 84% of the total variance.

The countries were subdivided into five groups (Fig. 3), where the countries with the best safety performance (first group) are Sweden, Norway, Switzerland, Great Britain, France, the Netherlands and Germany; the second group includes Finland, Denmark, Malta, Austria, Luxembourg and Ireland; and so on. Malta, Luxembourg and, especially, Italy, are 'outliers' of their groups, having much lower values of Factor 1 in comparison with the countries with a similar level of the combined safety indicator. (In this model, similar to the 'PCA-all', better safety position is associated with lower WF and Factor 1 values.)

3.1.4. FA with two factors

In the FA with two factors, two alternatives were considered: excluding and including the C4 indicator (average EuroNCAP score). Those were termed 'FA-2factors-noC4' and 'FA-2factors', respectively, and provided 71% and 69% of the explained variance.

In both trials, Factor 1 reflected mostly the behaviour of B1–B5 indicators (final outcomes), 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 reflected mostly the behaviour of policy performance indicators A1–A5, C2–C3 (safety belt use) and C1 (share of drink-driving accidents).

In both cases, similar behaviour was observed for basic indicators: B1-B2-B3-B5-C5 (final outcomes and median age of cars), A1-A2-A3-A4-A5 (policy performance), C2–C3 (use of seat belts) and D1–D2 (background characteristics). Moreover, C4 indicator is relatively close to the group of 'final outcomes' plus 'median age of cars', whereas C1 (share of drink-driving accidents) behaves

Fig. 2. Countries' positions on the dimensions of the combined factors (PCA-group analysis).

Fig. 3. Countries plotted using the composite indicator (WF) and Factor 1 values (FA-4factors analysis).

differently from all the other indicators. Besides, B4 (injury accidents per fatality) and C6 (percentage of motorcycles in fleet) are relatively close to the D1–D2-group.

In both cases the groups of countries with similar values of combined indicators were as shown in Fig. 4, where the countries with the best safety level (first group) are Malta, the Netherlands, Sweden, Great Britain, Germany, Switzerland, France and Norway; the second group includes Finland, Denmark, Austria, Luxembourg and Ireland; and so on. Similar to the results of the 'FA-4factors' analysis, Malta, Italy and Luxembourg behave as 'outliers' of their groups, and better safety position of a country is associated with lower values of WF and Factor 1.

3.2. Discussion: comparisons among the trials

The discussion is focused on two major issues. The first is an analysis of the consistency of the results of the various trials and identification of final groups of countries with similar safety performance – Sections 3.2.1 and 3.2.2. The second issue relates to the summing up of the behaviour of the basic indicators in the various trials, to recognize the existence of a stable group of indicators

Fig. 4. Countries plotted using the composite indicator (WF) and Factor 1 values (FA-2factors analysis).

– Section 3.2.3. Both issues are essential for further development of a composite road safety indicator.

3.2.1. Comparisons of countries' rankings

The composite indicators obtained for each country with each analysis (WF values) were compared graphically in Fig. 5. The values obtained for each country were plotted against the countries' ranking according to their final scores from the 'PCA-all'.

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, a certain similarity exists between the results of 'PCA-all', 'FA-4factors', 'FA-2factors' and 'FA-2factors-noC4' analyses.

The different character of the results obtained with the 'PCAgroups' method probably stems from the different approach taken in this case, where the indicators were initially analysed in the predefined groups, not accounting for the inter-group correlations. Such a consideration enables a deeper insight into the behaviour of indicators in each group, in comparison to other analyses, but can provide a different final picture when a combined indicator is created.

Accounting for their character, it was interesting to compare the intermediate results of the 'PCA-groups' analysis, i.e. the factors built for the A-, B-, C-groups of indicators, with the results of other analyses. Due to scaling reasons, prior to this step, we passed from the WF values to country ranks, where a country with the best value of the combined indicator receives rank '1', the next country – rank '2' and so on.

For example, Fig. 6 presents a comparison of country ranks according to the 'PCA-all' analysis and the group factor analyses: FB (final B-group factor), FC (final C-group factor) and FB1 – Factor 1 fitted in B-group analysis, which reflects five B-indicators together (fatalities per population, vehicles, passenger km-travelled as well as injury accidents per fatality and share of pedestrian fatalities). The basic countries' ranking is built using FB1 results, which actually represents a substitute to the traditional approach of countries' comparison in terms of fatality rates (per population, vehicles, etc.). To note, a similar comparison was performed by Hermans et al. (2008), who used the countries' ranking based on personal risk as a basis for judging the results of other rankings. The FB1 ranks in Fig. 6 are accompanied by ±5 rank deviation bars (vertical grey lines).

It can be seen from Fig. 6 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 population, vehicles, passenger kilometers travelled, etc.) in countries' estimation by 'final safety outcomes'.
Both 'PCA-all' ranks and FC ranks create deviations from the countries' ranks based on final 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.

Fig. 5. Final composite safety indicators (WF values) as a result of five analyses.

Fig. 6. Countries' ranks resulting from the FB1 ranks versus PCA-all, FC and FB ranks.

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) – Fig. 7. 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.

Fig. 7 shows that:

Similar to countries' ranking based on the combined indicators WF (see Fig. 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.

In general, the level of consistency between the different rankings seems average, whereas the most similar results were provided by a similar type of analysis (e.g. the FA variations).

3.2.2. Identification of groups of countries

Due to various reasons (e.g. data quality, method of analysis, random variations in data), the country rankings from different analyses will probably never be identical. Therefore, a more reasonable comparison of the results of several analyses could be made by forming groups of countries with similar safety performance. Such a comparison is possible in our case, because each analysis produced clusters of countries with very similar values of the composite indicator.

Using the results of countries' clustering observed in separate trials, the countries' classifications into five groups were compared. Table 6 summarizes the findings. The left part of Table 6 illustrates the country groupings obtained from the four analyses with more consistent results, i.e. 'FA-4factors', 'FA-2factors', 'FA-2factorsnoC4' and 'PCA-all', whereas the right part of the table illustrates the country groupings obtained from the initial and final steps of the 'PCA-groups' analysis. The two central columns give the final classifications of the countries.

Table 6 clearly shows that the countries' classifications in the left part are much more consistent than those in the right part of the table. Based on the left part of the table (results of four analyses), five groups of countries with different levels of safety performance can be defined as follows:

1. Countries with the 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.

Fig. 7. Final countries' ranks resulting from four analyses.

Table 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.

3.2.3. Behaviour of basic indicators

In the analyses performed, similarities in behaviour were observed for the following groups of indicators:

B1-B2-B3-B5-C5 (final safety outcomes – the fatality rates per population, per vehicles, per passenger kilometers travelled and share of pedestrian fatalities, and the median age of passenger cars),

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

C2–C3 (wearing rates of safety belts in front and rear seats) and D1–D2 (background characteristics).

Moreover, the C4 indicator (average EuroNCAP score) was relatively close to the group containing final 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 differently from all the other indicators, whereas indicators B6 (share of bicyclists fatalities), B7 (share of motorcyclist fatalities) and C7 (percentage of HGV in fleet) were excluded from several analyses due to low correlation with other variables.

We should mention here that the values of the indicators relating to safety policy have a rather arbitrary and subjective element, which could partly explain the homogeneity of their behaviour. As far as possible, the indicators were based on quantifiable facts, collected from published country reports. However, at this stage it is difficult to objectively quantify elements of a safety policy, and further research is required.

Using the scoring coefficients and factors' weights produced by each of the five analyses (see example in App. B), final weights of the basic indicators can be estimated and compared among the trials. Such weights may illustrate the contribution of various indicators to the composite safety indicator. Table 7 presents these final weights, which for a comparison are given in a normalized form (a sum in column is equal to 1). It can be seen that among the basic indicators considered (except for B6, B7 and C7, which were excluded from several analyses):

The largest differences between the trials are observed for B2 (fatalities per vehicles) and both background characteristics.
On average, higher weights were observed for B2, C1–C5 (intermediate outcomes), D2 (population density) and A2–A5 (policy indicators), where the lowest weights are associated with B4 (injury accidents per fatality) and B5 (share of pedestrian fatalities).
More consistent weights between the various trials are found for C1–C5 and A1–A5 groups of indicators.

However, as to the contribution to the composite indicator, the interpretation of the weights is not straightforward, since for the estimation of composite indicator for each country the original

Table 7 Final weights of the basic indicators based on the results of five analyses.

Basic indicator	PCA-all	PCA-groups	FA-4factors	FA-2factors-noC4	FA-2factors	Average value	Standard deviation
B1	0.091	0.039	0.039	0.081	0.013	0.021	0.063
B2	0.122	0.049	0.897	0.531	0.555	0.411	0.376
B3	0.166	0.047	0.107	0.133	0.008	0.070	0.093
B4	0.022	0.042	0.035	0.061	0.008	0.013	0.040
B5	0.089	0.042	0.003	0.039	0.033	0.011	0.054
B6	0.239	0.026	n/a	n/a	n/a	0.107	0.187
B7	0.209	0.049	n/a	n/a	n/a	0.129	0.113
C1	0.191	0.211	0.047	0.003	0.123	0.115	0.090
C2	0.243	0.157	0.417	0.197	0.111	0.225	0.118
C3	0.256	0.139	0.250	0.221	0.159	0.205	0.053
C4	0.155	0.073	0.155	n/a	0.063	0.111	0.051
C5	0.216	0.023	0.160	0.171	0.013	0.117	0.093
C6	0.257	n/a	0.011	0.005	0.075	0.079	0.125
C7	0.120	n/a	n/a	n/a	n/a	0.120	n/a
D1	0.181	0.479	0.035	0.012	0.082	0.080	0.244
D2	0.116	0.479	0.002	0.046	0.062	0.123	0.208
A1	0.016	0.135	0.147	0.099	0.062	0.092	0.054
A2	0.021	0.154	0.244	0.195	0.196	0.162	0.085
A3	0.004	0.131	0.194	0.075	0.130	0.107	0.071
A4	0.270	0.140	0.008	0.067	0.107	0.115	0.103
A5	0.113	0.153	0.173	0.152	0.195	0.157	0.031

Table A1 The final set of basic safety indicators for the 27 countries.

Country	A1	A2	A3	A4	A5	B1	B2	B3	B4	B5	B6	B7	C1	C2	C3	C4	C5	C6	C7	D1	D2
AT	1	3	2	2	1	88.1	174.6	99.7	54.6	0.15	0.07	0.18	0.06	83	52	20.5	7.5	0.12	0.07	507	99
BE	2	2	2	2	2	101.3	216.1	96.2	38.5	0.11	0.09	0.16	0.04	71	24	21.3	6.7	0.06	0.11	470	347
CY	1	2	2	2	1	111.3	236.2	167.5	31.1	0.22	0.02	0.29	0.23	80	36	20.7	9.3	0.08	0.22	479	84
CZ	2	3	2	2	2	103.5	263.5	145.8	20.8	0.19	0.10	0.11	0.06	72	13	20.7	11.0	0.15	0.09	399	130
DK	1	2	2	2	2	56.3	153.6	55.8	17.7	0.20	0.10	0.15	0.23	85	63	20.7	7.3	0.07	0.19	371	126
EE	2	3	2	2	1	151.8	389.4	175.3	12.7	0.27	0.04	0.04	0.28	74	30	21.7	12.7	0.02	0.14	413	30
FI	1	2	2	2	2	63.8	136.1	53.0	20.1	0.12	0.11	0.09	0.23	88	78	20.5	9.1	0.10	0.12	475	16
FR	1	1	2	1	1	76.8	153.1	63.9	17.1	0.11	0.04	0.23	0.29	97	70	20.7	8.3	0.06	0.14	504	113
DE	1	2	2	1	2	61.8	109.9	57.3	64.4	0.14	0.10	0.18	0.05	96	89	20.1	6.0	0.10	0.05	566	231
UK	1	2	2	1	1	54.4	115.7	47.6	58.9	0.21	0.05	0.18	0.17	90	84	20.6	5.0	0.04	0.11	471	249
EL	1	2	2	2	2	148.6	374.6	174.4	9.8	0.16	0.01	0.30	0.11	79	28	20.7	8.0	0.17	0.17	407	85
HU	1	3	2	2	2	129.0	446.0	271.2	16.1	0.23	0.12	0.10	0.09	67	34	20.6	8.0	0.04	0.13	293	108
IE	2	2	2	1	2	86.3	211.1	129.6	18.6	0.19	0.03	0.16	0.15	86	46	20.6	5.1	0.02	0.15	418	61
IT	3	4	3	3	4	96.2	162.1	73.8	42.0	0.13	0.05	0.24	0.02	71	21	21.0	7.9	0.20	0.09	597	196
LV	2	3	2	2	2	177.9	520.3	256.5	10.6	0.38	0.06	0.04	0.22	77	23	21.3	10.0	0.04	0.12	360	35
LT	2	3	2	2	2	223.6	498.1	190.6	8.9	0.34	0.05	0.12	0.12	60	13	22.6	14.5	0.01	0.08	470	52
LU	1	4	2	2	2	76.2	115.8	54.1	21.2	0.10	0.02	0	0	80	60	20.8	4.4	0.10	0.08	661	183
MT	1	3	2	2	2	24.6	46.4	48.8	84.1	0.35	0	0.18	0.25	96	28	18.6	9.0	0.04	0.16	535	1360
NL	1	1	2	1	1	44.7	101.9	48.4	33.6	0.09	0.18	0.18	0.15	90	64	20.2	6.0	0.06	0.11	442	394
NO	2	1	1	2	2	52.2	116.0	62.8	32.7	0.14	0.03	0.15	0.31	91	83	21.0	8.0	0.09	0.17	445	14
PL	2	3	2	2	2	137.5	407.6	234.9	8.9	0.32	0.11	0.04	0.10	78	23	20.6	11.8	0.05	0.14	351	122
PT	2	3	2	2	2	91.6	228.3	130.8	36.8	0.16	0.04	0.24	0.11	86	45	20.7	9.7	0.09	0.21	405	115
SK	2	3	2	2	2	107.4	439.1	215.0	13.8	0.09	0.06	0.07	0.12	76	24	21.0	7.0	0.04	0.12	247	111
SI	2	3	2	2	2	130.6	270.0	112.7	44.4	0.14	0.05	0.21	0.32	87	30	21.1	5.0	0.05	0.06	488	99
ES	1	3	2	2	2	93.0	200.7	117.2	24.3	0.15	0.02	0.18	0.12	74	51	20.6	7.6	0.15	0.17	464	88
SE	1	1	1	1	1	49.0	106.5	44.7	40.7	0.12	0.06	0.16	0.34	92	73	22.6	7.0	0.10	0.09	461	20
CH	1	1	1	2	1	49.6	95.0	42.1	58.1	0.21	0.09	0.22	0.19	82	53	20.5	7.3	0.12	0.06	519	182

Note: Prior to the calculations, the nominal values of the A-group indicators were converted into ordinal scale values. In our calculations we used the usual Pearson correlation for all the measures. This correlation treats all variable as scale variables. Using the ordinal measures as scales is equivalent in our case to the idea of using rank correlation. indicators should be standardized, i.e. mean centred and divided by their standard deviation (see Section 2.4), which may influence the sign and the value of the contribution.

Based on all the analyses performed in the study, basic safety indicators with consistent behaviour and a clearer contribution to the final composite safety indicator can be suggested. Such indicators are: B1-B2-B3 (fatality rates), A1-A2-A3-A4-A5 (quality of national safety programmes), 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.

4. Conclusions

In this study, we examined Principal Component Analysis and Common Factor Analysis weighting to create a composite safety indicator, based on four groups of basic road safety indicators of countries. The basic indicators considered referred to the three types of road safety benchmarking: policy performance indicators, final road safety outcomes and intermediate outcomes, and also included background characteristics of the countries (motorization level, population density). The analysis used the data collected for 27 European countries. Five trials of creating a composite indicator were performed, where each trial produced a composite road safety indicator for each country, and groups of countries with similar values of the composite indicator.

The purpose of our analysis was to explore the ways for creating a composite road safety indicator and the behaviour of basic indicators involved. It has been demonstrated that both tasks can be realized by the statistical weighting methods applied. The composite indicators, estimated by several methods, enabled us to rank the countries according to their safety performance.

Similar to other studies (Hermans et al., 2008) differences were observed in the countries' rankings obtained by the different methods used. It appears that a more reasonable comparison can be made by using groups of countries with similar levels of safety performance, where the countries are grouped based on the values of composite indicators received. In particular, among the 27 European countries considered, five groups with different safety performance were recognized, although, of course, a particular country could move from one group to another (neighbouring) group, depending on the method used. 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.

The analysis revealed that the results of countries' rankings based on the composite indicators are not necessarily identical to the traditional ranking they receive based on fatality rates only. Moreover, it was observed that the indicators belonging to the final outcomes (fatality and injury rates) and intermediate outcomes (wearing rates of seat belts, crashworthiness and composition of vehicle fleet, share of drink-driving) are not uniform in their behaviour in the analyses. Indicators which were found to be more consistent and influential and termed the 'core set of basic indicators' are recommended for future use.

The overall conclusion is that it is realistic and meaningful to design a composite road safety indicator in which relevant information from the different components of the road safety pyramid has been captured and weighted. Such an indicator gives a more

Table B2 Standardized scoring coefficients.

Basic indicator (variables)	Factor1	Factor2	Factor3	Factor4
B3	0.283	0.039	0.474	0.046
B2	0.832	0.191	0.348	0.202
B1	0.433	0.071	0.970	0.030
B5	0.017	0.050	0.055	0.011
C5	0.091	0.060	0.046	0.103
C2	0.085	0.045	0.332	0.203
C3	0.061	0.163	0.087	0.090
B4	0.049	0.051	0.041	0.102
D1	0.107	0.028	0.124	0.133
A4	0.041	0.371	0.240	0.282
A5	0.027	0.330	0.136	0.056
A2	0.122	0.143	0.263	0.168
C6	0.063	0.223	0.089	0.146
A1	0.088	0.039	0.344	0.013
C4	0.148	0.010	0.219	0.160
D2	0.109	0.180	0.143	0.413
A3	0.133	0.042	0.281	0.243
C1	0.030	0.044	0.149	0.241

Table B3 Factors' weights.

w1	w2	w3	w4
0.388	0.273	0.178	0.161

Table B1 Rotated factor pattern.

	Basic indicator and its ''safety-desired" behaviour	Factor1	Factor2	Factor3	Factor4
B3	;	0.931	0.162	0.164	0.099
B2	;	0.901	0.158	0.343	0.072
B1	;	0.686	0.247	0.594	0.114
B5	;	0.604	0.039	0.454	0.064
C5	;	0.527	0.208	0.463	0.235
C2	"	0.501	0.490	0.433	0.288
C3	"	0.588	0.472	0.412	0.255
B4	;	0.704	0.170	0.094	0.140
D1	"	0.821	0.139	0.195	0.174
A4	;	0.187	0.850	0.139	0.019
A5	;	0.028	0.819	0.008	0.196
A2	;	0.318	0.656	0.221	0.414
C6	;	0.456	0.587	0.307	0.055
A1	;	0.160	0.560	0.446	0.108
C4	"	0.096	0.002	0.773	0.325
D2	;	0.361	0.193	0.165	0.715
A3	;	0.096	0.482	0.013	0.659
C1	;	0.024	0.401	0.131	0.739

Table B4 Means and standard deviations of variables.

	B1	B2	B3	B4	B5	C1	C2	C3	C4	C5	C6
Mean Standard deviation	98.6 44.1	240.1 135.3	120.0 71.1	29.1 16.9	0.180 0.077	0.158 0.098	81.2 9.1	46.6 23.5	20.9 0.6	8.08 2.4	0.082 0.050
	N_A1	N_A2	N_A3	N_A4	N_A5	D1	D2
Mean Standard deviation	1.530 0.566	2.381 0.871	1.937 0.405	1.795 0.489	1.784 0.638	449.3 88.0	126.6 95.6

enriched picture of road safety than a ranking based only on data on fatality rates, which is normal practice today. Grouping countries in this process is promising and seems to be preferable to simply ranking countries. Before defining the composite road safety indicator and actually using the results for policy making, two improvements are recommended: firstly, to design the final set of basic safety performance indicators and, secondly, to develop procedures for collecting high-quality and comparable data for the indicators for all the countries compared.

Acknowledgements

The authors would like to acknowledge the SUNflowerNext team (Wegman et al., 2008) for the inspiration, multiple fruitful discussions and assistance in data collection for the study's performance.

Appendix A

See Table A1.

Appendix B

Results of the 'FA-4factors' analysis See Tables B1–B4.

References

Al Haji, G., 2005. Towards a road safety development index (RSDI). Development of an international index to measure road safety performance. Linkoping studies in Science and Technology, Licentiate Thesis No. 1174. Department of Science and Technology, Linkoping University.
Broughton, J. Buckle, G, 2006. Monitoring progress towards the 2010 casualty reduction target – 2004 data. TRL Report 653. TRL Limited, Wokingham.
European Commission, 2007. Energy and Transport in Figures. Part 3: Transport. Directorate-General for Energy and Transport in co-operation with Eurostat. (Chapter: 3.6 Means of Transport; Chapter 3.7 Safety).
Elvik, R., 2001. Improving road safety in Norway and Sweden: analyzing the efficiency of policy priorities. Traffic Engineering and Control January, 9–16.
Elvik, R., 2008. Road safety management by objectives: a critical analysis of the Norwegian approach. Accident Analysis and Prevention 40, 1115–1122.
Elvik, R., Amundsen, A.H., 2000. Improving road safety in Sweden. Report 490. Institute of Transport Economics, Oslo.
European Road Safety Observatory, 2008. Annual statistical Report 2007. <http:// www.erso.eu>
European Transport Safety Council, 2001. Transport Safety Performance Indicators, Brussels.
European Transport Safety Council, 2006. A Methodological Approach to National Road Safety Policies, Brussels.
European Transport Safety Council, 2007a. Reducing Deaths From Drink Driving. PIN Flash 5 Background data.
European Transport Safety Council, 2007b. Increasing Seat Belt Use. PIN Flash 4 Background data.
Hakkert, A.S., Gitelman, V., Vis, M.A. (Eds.), 2007. Road Safety Performance Indicators: Theory. Deliverable D3.6 of the EU FP6 Project SafetyNet.
Hermans, E., Van der Bossche, F., Wets, G., 2008. Combining road safety information in a performance index. Accident Analysis and Prevention 40, 1337–1344.
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. Transforming variables for normality and linearity When, How, Why and Why Not's. In: SAS Conference Proceedings NESUG 2005, 11–14 September, Portland, Maine.
Nardo, M., Saisana, M., Saltelli, A., Tarantola, S., 2005. Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Statistics Working Papers, STD/DOC(2005)3. Organisation for Economic Co-operation and Development.
Organisation for Economic Co-operation and Development, 2002. Safety on Roads What's the Vision, Paris.
Organisation for Economic Co-operation and Development/European Conference of Ministers of Transport, 2008. Country Reports on Road Safety Performance. OECD/ECMT working group on achieving ambitious road safety targets, Joint OECD/ECMT Transport Research Centre, Paris. <http://www. internationaltransportforum.org/jtrc/safety/targets/Performance/performance. html>.
Organisation for Economic Co-operation and Development/International Transport Forum, 2008. Towards Zero – Ambitious Road Safety Targets and the Safe System Approach. Joint Transport Research Centre of OECD/ITF, Paris.
Peden, M., Scurfield, R., Sleet, D., Mohan, D., Hyder, A., Jarawan, E., Mathers, C. (Eds.), 2004. World Report on Road Traffic Injury Prevention. World Health Organisation, Geneva.
Tabachnick, B.G., Fidell, L.S., 2006. Using Multivariate Statistics. Pearson/Allyn & Bacon, Boston.
UNECE Database, 2008. <http://w3.unece.org/pxweb/Dialog/varval.asp?ma=01\_ TRRoadtypVeh\_r&ti=Road+vehicle+fleet+at+31+December+by+Type+of+vehicle% 2C+Age+group%2C+Country+and+Year&path=./DATABASE/Stat/40-TRTRANS/02- TRRoadFleet/&lang=1>.
VESIPO, 2002. Erarbeitung der Grundlagen fur die Strassenverkehrssicherheitspolitik des Bundes u.a. Basler & Hofman, Eckhardt, Perrin, Schonenberger, Fierz.
Vis, M.A., Van Gent, A.L. (Eds.), 2007. Road Safety Performance Indicators: Country Comparisons. Deliverable D3.7a of the EU FP6 Project SafetyNet.
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.
Wegman, F., Commandeur, J., Doveh, E., Eksler, V., Gitelman, V., Hakkert, S., Lynam, D., Oppe, S., 2008. SUNflowerNext: Towards a Composite Road Safety Performance Index. SWOV Institute for Road Safety Research, Leidschendam, the Netherlands.

gitelmandesigningcompositeindicatorroad2010.md 123 KB History Raw