Assessing Road Safety Development in European Countries: A Cross-Year Comparative Analysis of a Safety Performance Index ^†

Abstract

The development of a road safety performance index has been widely accepted as a supportive instrument to evaluate and compare the safety performance in different countries. However, most of the current studies concentrate on the index development for only one year. In other words, there is still a lack of cross-year comparison based on the constructed safety performance index, so as to assess the progress of road safety performance in different countries over time. In this study, by collecting data on four background indicators and seven safety performance indicators of 21 European countries for both 2008 and 2015, the hierarchical clustering analysis is first utilized to identify country groups based on the background indicators. Then, the principal component analysis (PCA) is applied for each group to construct a safety performance index, and a cross-year comparative analysis on country grouping, index ranking, and weight allocation is conducted. The results show that the members in the two country groups remain the same, implying that there was no dramatic change with respect to these countries’ road safety policy context in these two years. However, the gaps between these two country groups with respect to their overall socioeconomic development, as well as their road safety performance, enlarged over this period. Moreover, by comparing the indicator weights assigned for each country group in different years and examining the changes in indicator values of each country, a useful insight into the areas of underperformance is gained, which cannot be revealed in single year index evaluation. All these findings provide policy makers with valuable guidance to prioritize their actions to improve the level of road safety.

1. Introduction

Road safety is closely related to the social and health condition of human beings and to the sustainable development of every country [1]. The World Health Organization (WHO) reported that around 1.35 million people die due to road crashes each year around the world. Road traffic injuries are the eighth leading cause of death for people of all ages and the first cause of death for children and young adults aged 5–29 years [2]. They not only impose large economic costs, but also cause great emotional stress to those affected families [3]. Hence, it is of great importance to make continuous progress in road safety around the world.

Having recognized the importance of the problem and the need for action, the United Nations launched the second “Decade of Action for Road Safety 2021–2030” with the explicit target of reducing road fatalities and injuries by at least 50% during this period (UN General Assembly Resolution 74/299). For this purpose, apart from recording and analyzing the number of crashes and casualties (see, e.g., [4,5,6,7]), various underlying risk factors that contribute to crashes should be studied in detail. Therefore, investigation on safety performance indicators (SPIs) is widely encouraged [8]. They are defined as any measurement that is causally related to the number of crashes or to the injury consequences of a crash (e.g., the seat belt wearing rate of car occupants), and are used in addition to the figures of crashes or injuries to understand the process that leads to crashes [9].

Today, having been aware of the complex character of the road safety issue, a great number of SPIs have been developed [10,11,12,13,14,15,16,17]. Hence, combining various SPIs into an index is regarded as a scientifically sound way to reduce the dimension of selected SPIs. Such an index can then be used as a supportive instrument to compare and prioritize road safety actions among various countries. Some of the successful examples are presented below. Hermans et al. explored a methodological framework for creating a road safety performance index for cross-country comparison, in which seven SPIs with respect to alcohol, speed, protective systems, daytime running lights, vehicle, infrastructure, and trauma care were defined, and five weighting approaches were applied to create one overall index for 21 European countries. The study concluded that comparing the road safety performance of countries by means of an index enabled earlier and goal-oriented action [10]. This study was further extended by Bao et al. [11], in which some additional SPIs were considered, constituting a layered hierarchy, and the technique of hierarchical fuzzy TOPSIS was developed for the construction of an overall index. In the SUNflowerNext study [12,13], three types of performance indicators, i.e., road safety performance indicators, implementation performance indicators, and policy performance indicators, were distinguished, and an overall road safety index was developed by two weighting approaches (i.e., principal component analysis and factor analysis) based on the data collected for 27 European countries. The analysis revealed that the countries’ ranking based on the combination of different performance indicators may be different from the traditional ranking of countries based on crash data. This study was revisited by Shen [14], in which the different treatment between quantitative and qualitative indicator data is considered when developing such an index. In the DaCoTA study [15], a two-dimensional index was created, where countries were ranked simultaneously on the basis of the final outcome index and the safety performance index. The aforementioned studies clearly demonstrated the feasibility and usefulness of developing an overall index for road safety performance evaluation among a group of countries. However, all these studies concentrated on the index development for only one year. In other words, there is still a lack of cross-year comparison based on the constructed safety performance index [18].

In fact, by collecting the SPI data and computing the corresponding index score at regular intervals, road safety policymakers can systematically evaluate the results of various policy interventions, monitor the progress of road safety performance in each country, and make meaningful comparisons among different countries [19]. Hence, the current research gap motivated us to focus on the comparison of road safety performance for a number of countries between different years based on the developed index. To this end, comparable countries are first grouped, as it is more reasonable for countries to learn from similar ones than from those differing in physical and social-economic characteristics. More specifically, 21 European countries are considered in this study, and they are divided into two groups by using the hierarchical clustering analysis. After that, seven quantitative SPIs that currently have data available for all these countries are selected, and corresponding data for both 2008 and 2015 are collected. The principal component analysis (PCA) approach is then applied to combine these SPIs into an overall index for these two years, respectively. Cross-year comparisons on country grouping, index ranking, and weight allocation are conducted afterwards to assess the road safety development of these countries during this period, identify areas of their underperformance, and formulate road safety priorities for policy action.

The remaining part of this paper is organized as follows. In Section 2, the indicators used and the data collected for this study are briefly introduced. Section 3 focuses on the methodology for country grouping and safety performance index construction. The results on cross-year comparison are presented and discussed in Section 4. Finally, the paper ends with main concluding remarks and directions for future research in Section 5.

2. Indicator Selection and Data Collection

In this study, two types of road safety related indicators are considered. One is background indicators, which are used to divide countries into comparable groups, and the other is safety performance indicators, which are used to construct the road safety performance index.

The background indicators give an essential background condition of a country or its policy context. Progress in road safety may not be fully understood or even be misinterpreted by not knowing or ignoring this background information [9]. The background indicators should influence the road safety level of countries and hardly be directly influenced by the road safety policy of countries [12]. According to the above characteristics and taking data availability and reliability into account, four background indicators describing the level of a country’s motorization, economic development, population density, and urbanization are chosen and shown in

Table 1. Selected background indicators for country grouping.

Background Characteristics	Selected Background Indicators	Abbreviation
Motorization	The number of passenger cars per 1000 inhabitants	N_p_cars
Economic development	Gross Domestic Product (GDP) per head, EU27 = 100	GDP
Population density	Population per 1 km2 of country’s territory	Pop_den
Urbanization	The percentage of population living in urban areas	Pop_urb

. They are (in parentheses are abbreviations): the number of passenger cars per 1000 inhabitants (N_p_cars); Gross Domestic Product (GDP) per head (GDP); population per 1 km² of a country’s territory (Pop_den); the percentage of population living in urban areas (Pop_urb).

The safety performance indicators are the measures reflecting those operational conditions of the road traffic system that influence the system’s safety performance [20]. They can show in more detail the state of risk factors and their trends, as well as the potential to reduce these types of crashes [21]. Ideally, SPIs should encompass the most important risk areas related to humans, vehicles, and roads. Based on the literature review, seven SPIs with currently available data for two different years are selected, which represent three risk domains, i.e., alcohol, protective systems, and vehicles (see

Table 2. Selected safety performance indicators for index construction.

Risk Domain	Selected SPIs	Abbreviation
Alcohol	I1: The percentage of drivers below alcohol limit in roadside checks	P_alc
Protective systems	I2: Daytime seat belt wearing rates on front seats	Belt_front
	I3: Daytime seat belt wearing rates on rear seats	Belt_rear
Vehicles	I4: Average score of occupant protection for new cars	Prot_occ
	I5: Average score of pedestrian protection for new cars	Prot_ped
	I6: Renewal rate of passenger cars	Renewal
	I7: The percentage of new passenger cars: less than 5 years	Car_age

). Specifically, the percentage of drivers complying with the alcohol limit in roadside checks is the selected alcohol indicator (P_alc); the daytime seat belt wearing rates on front seats and rear seats of passenger cars are chosen for the domain of protective systems (Belt_front and Belt_rear); and the vehicle domain is represented by the average score of occupant protection for new cars (Prot_occ), the average score of pedestrian protection for new cars (Prot_ped), the renewal rate of passenger cars (Renewal), and the percentage of new passenger cars of a maximum of five years old (Car_age). The first two indicators in this domain are generated from the European New Car Assessment Programme, i.e., EuroNCAP, where vehicle crash performance is evaluated by rating the vehicle models according to their safety level for occupant protection (1 to 5 stars), pedestrian protection (1 to 3 stars), etc. [22].

From several international databases and publications [22,23,24,25,26], indicator values related to both 2008 and 2015 are collected for 21 European countries. They are Austria (AT), Belgium (BE), Switzerland (CH), Czech Republic (CZ), Germany (DE), Denmark (DK), Estonia (EE), Spain (ES), Finland (FI), France (FR), Hungary (HU), Ireland (IE), Italy (IT), Lithuania (LT), the Netherlands (NL), Poland (PL), Portugal (PT), Romania (RO), Sweden (SE), Slovakia (SK), and the United Kingdom (UK). The missing values are imputed by using the Expected Maximization method in SPSS 26.0 [27]. The indicator values are presented in Appendix A.

3. Methodology

3.1. Clustering Analysis for Country Grouping

Before combining the selected SPIs in an index, the 21 European countries considered in this study are first classified into several groups according to their background characteristics to make the following comparison more logical and meaningful. Moreover, those underperforming countries might be more motivated to improve themselves if being the ‘best-in-class’ is considered to be within reach [15].

The subdivision of these countries into homogeneous groups is realized by applying a clustering analysis (CA). The main purpose is to divide observations in a dataset into subsets, where objects in the same subset are similar to each other with respect to a certain similarity measure [28]. Over the past decades, a large number of clustering methods such as hierarchical clustering, K-means clustering, spectral clustering, etc., have been developed and successfully applied in a wide range of fields, such as biology, geography, climate, psychology, medicine, and business (see e.g., [29,30]).

The clustering is normally based on the standardized data, where the raw data are first normalized by converting them into standard scores, as follows:

$$z_i=\frac{x_i-\overline{x}}{sd}$$

(1)

where $\overline{x}$ denotes the sample mean, and sd denotes the sample standard deviation.

Then, a distance measure should be selected and computed. A distance measure is an estimate of the degree of similarity between cases in the set. A small distance means a strong similarity. Some of the most commonly used distance measures include for example the Euclidean distances, the squared Euclidean distances, the Chebychev distances, and so on [31]. In this study, the squared Euclidean distance is chosen as a distance measure, because this measure is less affected by the addition of new objects such as outliers. It can be computed as follows:

$$D(m,n)=\frac{{\displaystyle \sum _{i=1}^{N_d}{(m_i-n_i)}^2}}{N_d}$$

(2)

where D(m, n) denotes the distance measures for individual indicators between two objects’ (countries in this study) m and n over N_d dimensions.

Bearing some basic ideas about the clustering analysis in mind, a certain clustering method can be applied for country grouping, e.g., the hierarchical clustering analysis, which can be employed directly in SPSS 26.0. Meanwhile, to test the validity of the clustering results, a hypothesis testing (for two groups) or ANOVA (for three or more groups) could be conducted.

3.2. Principal Component Analysis for Index Construction

Weighting and aggregation play an important role in the construction of an index. Various techniques are currently available in the literature to weight and aggregate different indicators for different contexts [31]. However, there is no best one to use in all circumstances. In this study, we do not concentrate on introducing new techniques for index construction, but on the comparison of index scores over time via an easy-to-reproduce but effective approach, which is the principal component analysis (PCA).

A PCA is a kind of multivariate data analysis method, and it is used frequently to reduce the dimensions of a problem. The main idea of this method is to generate a few so-called principal components (p) to replace a large number of original variables (l) without dropping important information. Several guidelines can be used to identify the optimal number of principal components. First, the reserved principal components should have associated eigenvalues larger than one; second, they should individually account for over 10% of the total variance; finally, they should cumulatively explain at least 60% of the overall variance [31].

Prior to the application of PCA for index construction, we should first normalize the raw data so that the scale differences of the indicators and the effects of the measurement unit can be eliminated. In this study, the distance between the actual value and the best and the worst values is calculated, resulting in a normalized value between 0 and 1, in which a value of 1 indicates the highest safety performance, while a value of 0 means the lowest.

After data normalization, we further obtain the matrix of component loading. Each indicator (i) can have a large loading score (a_ij (i = 1, …, l; j = 1, …, p)) on only one of the principal components (j). The indicator weights can then be derived based on these loading scores by the following procedures [10].

Define:

$$u_{ij}=\frac{a_{ij}^2}{{\displaystyle \sum _{i=1}^l{\displaystyle \sum _{j=1}^p{a}_{ij}^2}}}$$

(3)

Thus, the preliminary weight for each indicator i can be generated as follows:

$$u_i=\underset{j}{\max }\left(u_{ij}\right)$$

(4)

However, due to dimension reduction, the sum of u_i is less than one, i.e.,

$$U={\displaystyle \sum _i{u}_i}<1$$

(5)

In order to enable the sum of weights to be one, the final weight of indicator i is calculated by:

$$w_i=\frac{u_i}{U}$$

(6)

Based on the derived weights, the index score of country k can be obtained as follows:

$$Scor{e}_k={\displaystyle \sum _{i=1}^l{w}_i{v}_{ik}}$$

(7)

where v_ik is the normalized value of indicator i for country k.

4. Application and Results

In this study, based on the indicator data collected for the 21 European countries for both 2008 and 2015, the hierarchical clustering analysis is first utilized to identify country groups, and the principal component analysis is then applied for each group to calculate the weights of the selected SPIs and to obtain a safety performance index score for each country. In this section, the results on country grouping, index ranking, and weight allocation are presented and further compared between these two years.

4.1. Cross-Year Comparison of Country Grouping

By applying the hierarchical clustering analysis based on the four background indicators introduced in Section 2, different numbers of the clusters or country groups can be considered, and we select two for both 2008 and 2015, as in such a case each group can contain a reasonable number of countries for performance evaluation and cross-country comparison (see Figure 1 and Figure 2). If more clusters are considered, the number of countries in some clusters might be too small. For instance, Belgium and the Netherlands will form a separate group if one more cluster is considered based on the classification trees shown in Figure 1 and Figure 2.

It is interesting to notice that the results on country grouping from Figure 1 and Figure 2 show that the members in each cluster do not change over time. One group contains 13 countries (i.e., Austria, Belgium, Denmark, Finland, France, Germany, Ireland, Italy, The Netherlands, Spain, Sweden, Switzerland, and the United Kingdom), and the other contains the remaining eight countries (i.e., Czech Republic, Estonia, Hungary, Lithuania, Poland, Portugal, Romania, and Slovakia), for both 2008 and 2015.

The consistency of country grouping in these two years based on the same clustering approach facilitates our group-based country comparison. A colored international map of these 21 European countries is shown in Figure 3, which indicates the geographical distribution of these countries in relation to their background characteristics. It can be seen that most of the western and northern European countries (except Portugal) belong to Group 1, while most eastern European countries are within Group 2.

To further explore the differences between these two country groups, we calculate the average values of these four background indicators related to different country groups and different years and perform a t-test to measure the level of their differences. The results are shown in

Table 3. The average values of four background indicators related to different country groups and different years, as well as the p-value from the hypothesis testing.

		N_p_cars	GDP	Pop_den	Pop_urb
Year 2008
	Country Group 1	486	119	166	0.778
	Country Group 2	369	66	95	0.647
	p-value of t-test	0.0067	4.69 × 10−9	0.0320	0.0012
Year 2015
	Country Group 1	504	131	179	0.792
	Country Group 2	424	46	94	0.655
	p-value of t-test	0.0320	8.99 × 10−10	0.0284	0.0009

.

We can see that for both 2008 and 2015, the average value of each of the four background indicators of country Group 1 is statistically significantly larger than that of country Group 2 (with the p-value less than 0.05). Hence, from the view of motorization level, economic development, population density, and urbanization level, the countries in Group 1 are more developed than those in Group 2. Thus, country Group 1 can be treated as “Highly Developed Countries” and country Group 2 as “Less Developed Countries”.

Moreover, the development trend with respect to these four aspects was different for the two country groups between 2008 and 2015. As for Group 1, all indicators’ average values increased to a certain extent during these seven years. However, the situation was completely different for country Group 2. Except for N_p_cars, whose average value increased significantly from 369 passenger cars per 1000 inhabitants in 2008 to 424 in 2015, the other three indicators’ average values basically did not change (e.g., Pop_den and Pop_urb), or even decreased dramatically (e.g., the GDP per head dropped from 66 to 46) from 2008 to 2015. This implies that apart from the motorization level, the overall socioeconomic development gap between these two country groups enlarged during these seven years. Such a result can also be verified by the changing direction of the p-values from the t-test in Table 3.

4.2. Cross-Year Comparison of Country Ranking

By combining the seven SPIs via the principal component analysis approach, the overall safety performance index scores and corresponding rankings related to the two country groups are computed for the years 2008 and 2015, respectively. The results are shown in

Table 4. Index scores and corresponding rankings of countries within each group for different years.

Country	2008		2015
	Index Score	Ranking	Index Score	Ranking
Country Group 1
AT	0.579	11	0.705	6
BE	0.614	10	0.769	3
CH	0.532	12	0.688	8
DE	0.691	4	0.800	2
DK	0.629	9	0.575	10
ES	0.710	3	0.528	11
FI	0.664	5	0.719	5
FR	0.647	6	0.528	12
IE	0.865	1	0.746	4
IT	0.159	13	0.221	13
NL	0.631	8	0.636	9
SE	0.747	2	0.870	1
UK	0.632	7	0.702	7
Average	0.623		0.653
Country Group 2
CZ	0.667	4	0.804	1
EE	0.854	1	0.66	2
HU	0.748	2	0.394	7
LT	0.631	5	0.6	5
PL	0.478	7	0.607	4
PT	0.699	3	0.631	3
RO	0.242	8	0.196	8
SK	0.558	6	0.479	6
Average	0.610		0.546

.

As can be seen, Ireland obtains the highest index score of 0.865 and ranks the first among all the countries in Group 1 for the year 2008, but it drops to fourth place in 2015. Sweden moves from second place to the top over these seven years, with an increasing index score from 0.747 to 0.870. On the other hand, Italy keeps last place in this group, although its index score improves from 0.159 in 2008 to 0.221 in 2015. Regarding country Group 2, Estonia is the best-performing country for 2008 with an index score of 0.854, but such a ranking is replaced by the Czech Republic in 2015, with a somewhat lower index score of 0.804. Romania always ranks last within this group, and its index score also declines from 0.242 to 0.196. In general, from 2008 to 2015, nine out of 13 countries within Group 1 see their index score increase, and the overall road safety performance of the countries in this group improves with an average index score increased from 0.623 to 0.653. However, the situation is different for the countries in Group 2. Six out of eight countries obtain a lower index score, and the average index score of all these eight countries declines from 0.610 to 0.546, implying that the gap between these two country groups with respect to the road safety performance expanded during these seven years.

Regarding the ranking of these 21 European countries based on their index scores, we further use Figure 4 to illustrate its changes between these two years. As we can see, a relatively large positive change in ranking occurs for Belgium and Austria in country Group 1 and to the Czech Republic and Poland in country Group 2. In contrast, Spain and France in Group 1 and Hungary in Group 2 see a dramatically negative shift in their rankings from 2008 to 2015. Taking Belgium as an example, it achieved the largest improvement in ranking among all the countries in Group 1, moving from 10th place in 2008 to third place in 2015. Such a result can mainly be attributed to the uptrend change in most of its normalized SPI values (see Figure 5a). On the contrary, Hungary suffered the biggest retrogression within Group 2, moving from the second best-performing country in 2008 to the last but one in 2015. As can be seen from Figure 5b, all safety performance aspects of this country deteriorated over these seven years relative to the other countries within the same group, thus, resulting in this big decline in ranking.

In addition, the relationship between the developed safety performance index and the road safety final outcome indicator (i.e., the number of road fatalities per million inhabitants, see Appendix B) of these countries was also investigated. We found that in general, the countries in Group 1 had a fewer number of fatalities per inhabitants than the countries in Group 2 for both 2008 and 2015. Such a result is consistent with the results based on the developed safety performance index. However, the countries’ ranking based on the derived index score is different from the traditional ranking of countries based on the fatality data, which is in line with the findings from the SUNflowerNext study and the DaCoTA study [12,15]. Moreover, if we take the percentage change of the road fatalities per million inhabitants during these two years into account, the variance is large, which implies that we may not simply or completely attribute the percentage reduction in road fatalities of one country by its improvement in safety performance. In doing so, some other relevant aspects, such as the changes in socioeconomic conditions, in motorization levels, and in road safety policy interventions of each country might also be taken into account.

4.3. Cross-Year Comparison of Indicator Weights Allocation

After comparing the overall road safety performance of these 21 European countries in their own group between 2008 and 2015, we further investigate the indicator weights assigned for each country group in different years based on the principal component analysis approach. The results are shown in Figure 6.

From the view of time dimension, two out of seven SPIs’ weights increase from 2008 to 2015 for country Group 1, while the remaining weights show an opposite trend. Amongst others, the weight for Prot_ped increases by three times (from 0.065 to 0.195), and the weight for P_alc declines by more than half (from 0.166 to 0.075). The others only change slightly. Regarding country Group 2, there are also two indicators whose weights change dramatically in these two years. One is Belt_rear, which sees a large increase in the assigned weight (from 0.058 to 0.220); the other is P_alc, whose weight declines from 0.185 to 0.033. If we further compare the weight allocation between these two country groups, it can be seen that the largest difference in 2008 is Belt_rear (0.173 vs. 0.058), followed by Prot_ped (0.065 vs. 0.098). In 2015, the largest difference appears to be P_alc (0.075 vs. 0.033), also followed by Prot_ped (0.195 vs. 0.115). In general, P_alc, Belt_rear, and Prot_ped, which belong to the three different risk domains considered in this study (i.e., alcohol, protective systems, and vehicle, respectively), are the three SPIs that have the highest variation from the view of either time dimension or country grouping. Although it was indicated in the literature that weights derived from the PCA approach cannot be directly used as a measure of the theoretical importance of the associated indicator [31], given the data-driven nature of this approach, it is important evidence to show that road safety development with respect to these three indicators over this period varied considerably among these 21 European countries. Hence, more attention should be paid to these three aspects when formulating road safety priorities for policy action in these countries.

Taking Spain as an example, it suffered the biggest decline in ranking within country Group 1. Figure 7 illustrates the change in its normalized SPI values between 2008 and 2015. We can see that the values of its four vehicle-related SPIs declined largely over this period. Given the fact that the indicator weight assigned to Prot_ped for country Group 1 increases dramatically from 2008 to 2015, the low index score obtained by Spain for 2015 can then mainly be attributed to its poor performance on this indicator. Contrastingly, if more efforts were given to the improvement of pedestrian protection in Spain, such as installing infrared automatic brake control system in new cars, its overall index score as well as the corresponding ranking would improve accordingly.

It should be highlighted here that the decline in the normalized SPI values of a country does not always mean a retrogression of its actual performance, but a change relative to the other countries under consideration. It can be treated as an added value of using the normalized data for country evaluation and comparison. Looking at the change in the original values of the indicator Prot_ped over time (see Figure 8), all the countries in Group 1 made considerable progress in this respect from 2008 to 2015. However, relative to the other countries within the group, the improvement made by Spain was smaller, thus, resulting in a big drop in ranking with respect to this indicator and, consequently, a declined normalized value for 2015. Hence, Spain should not be satisfied with the progress it made during this period, but continue to prioritize policy in this aspect. Meanwhile, the progress made by the other countries in the same group also implies the applicability of this improvement direction.

5. Concluding Remarks

Over the past decade, a great number of safety performance indicators have been developed describing the complex character of the road safety phenomenon, and the construction of a safety performance index has been widely accepted as a supportive instrument to reduce the dimension of selected SPIs and to compare and prioritize road safety actions among various countries. However, due partly to the data unavailability, most of the current applications concentrated on the index development for only one year. In other words, there is a lack of cross-year comparison based on the constructed safety performance index, so as to assess the progress of road safety performance in different countries over time. In this study, by collecting the data on four background indicators and seven safety performance indicators of 21 European countries for both 2008 and 2015, a safety performance index was developed and a cross-year comparative analysis on country grouping, ranking, and weight allocation was conducted. The main results are summarized as follows:

• By employing the hierarchical clustering analysis on the four background indicators describing the level of a country’s motorization, economic development, population density, and urbanization, respectively, the 21 European countries were classified into two homogeneous groups, and the members in each country group remained the same in these two years, implying that there was no dramatic change with respect to these countries’ road safety policy context in 2008 and 2015. In general, most of the western and northern European countries were within one group and could be treated as “Highly Developed Countries”, while most eastern European countries belonged to the other group and could be treated as “Less Developed Countries”. However, the overall socioeconomic development gap between these two country groups enlarged during these seven years.
• By applying the principal component analysis approach to combine the seven SPIs into an index, the overall safety performance index scores and the corresponding rankings related to the two country groups were derived for 2008 and 2015, respectively. Belgium and Austria achieved a large improvement in ranking within country Group 1, and the Czech Republic and Poland in country Group 2 over this period. In contrast, Spain and France in Group 1 and Hungary in Group 2 suffered a large retrogression from 2008 to 2015. On the whole, the average index score increased for country Group 1 but declined for country Group 2, implying that the gap between these two country groups with respect to the road safety performance also expanded.
• Based on the indicator weights assigned for each country group in different years via the principal component analysis, P_alc, Belt_rear, and Prot_ped, which belong to the three different risk domains considered in this study (i.e., alcohol, protective systems, and vehicle), were found to be the three SPIs that had the highest variation from the view of either time dimension or country grouping, implying that the road safety development of these 21 European countries with respect to these three indicators differed considerably over this period, and thus, more attention should be paid to these three aspects when formulating road safety priorities for policy action in these countries.

Different from evaluating the road safety performance of a set of countries at one specific point of time by developing a safety performance index, a cross-year comparative analysis based on such an index was proven to be valuable in gaining a clear understanding of the road safety development of these countries over time. However, it should be noticed that the road safety problem is so complex that the development of mature road safety indicators and the collection of high-quality indicator data are still ongoing. In this study, only seven SPIs related to three risk domains were considered for index construction, and the indicator data for only two different years that are currently available were collected. In the future, other risk factors that have a strong relationship with road safety, such as speed violation, road conditions, etc., could be considered and corresponding indicators refined. Moreover, by collecting the SPIs data at regular intervals, systematic country comparison over time (such as every year) could be conducted so that the results of policy interventions would be evaluated and the progress towards postulated road safety targets would be monitored. In addition, regarding the weighting method adopted in this study, given the fact that the weighting strategy of PCA is to correct for overlapping information between those correlated indicators, other weighting methods such as analytic hierarchy process (AHP), data envelopment analysis (DEA), etc., are worthwhile to investigate in the future to measure the importance of the associated indicator on the one hand, and to embody the layered hierarchy of the indicators on the other. Finally, apart from investigating the safety performance index and its change over time, it is also valuable to compare it with the crash data in each country, so as to explore their potential causal relationship, together with other relevant influencing factors.