Road Safety Trends at National Level in Europe: A Review of Time-series Analysis Performed during the Period 2000–12

Abstract

This paper presents a review of time-series analysis of road safety trends, aggregated at a national level, which has been performed in the period 2000–12 and applied to European national data sets covering long time periods. It provides a guideline and set of best practices in the area of time-series modelling and identifies the latest methods and applications of national road safety trend analysis in Europe. The paper begins with the methodological framework adopted for aggregate time-series modelling that will be considered, and then discusses a number of relevant applications to long-period data aggregated at the national level, whether for countries alone, or for groups of countries. Some analyses, which were performed at the disaggregated level, are also provided, as they are being used more and more. Finally, the paper summarizes and discusses the significant changes in aggregate road safety trend analysis which occurred during the period and provides recommendations for continuing these research efforts.

Keywords:

• road safety
• time-series modelling
• trend analysis
• national level
• Europe

Acknowledgements

The authors would like to thank their colleagues from ICTSA, SafetyNet WP7 and DaCoTA WP4 networks, in particular Siem Oppe, Jacques Commandeur, Emmanuelle Dupont and Heike Martensen, for their collaboration in joint research efforts, their suggestions and comments. The authors would also like to thank Rosalind Greenstein for her careful rereading of the English.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. International Cooperation on Time Series Analysis, 2000–2006.

2. SafetyNet — Building the European Road Safety Observatory, WP7 — Data analysis and synthesis, 2004–2008.

3. DaCoTA — Road safety Data, Collection, Transfer and Analysis, WP4 — Decision Support, 2010–2012.

4. To describe the main differences between ARIMA and state space models (a more general denomination commonly used in the place of their particular case which is considered in this paper — namely the structural models) will help readers who are unfamiliar with these time-series techniques. When they are used in their simplest form without exogenous variable, both techniques aim to fit the observed data by referring to its past values only, which allows for easy and immediate forecasts of the variable of interest measured with the data. When exogenous variables are used, forecasts of the variable of interest are provided on the basis of scenarios for the development of the exogenous variables at the horizon of the forecast.

The main difference in these two techniques is in the model's output. While the ARIMA model fits the observed data as a whole, the state space model fits each of its unobserved components: the locally linear trend (its level and slope), the seasonal component and the residue. This special case of state space model is therefore called “structural” as it fits the hidden structure of the data. The file use the model for an application will lead the user to choose one or the other technique.

For the statistician, however, a relevant difference in these techniques is in the number of parameters. In case the data are fitted using state space methods — therefore with a stochastic approach — the variances of several parameters (the level and slope of the local linear trend, the s-1 dimension seasonal component and the final residue) are to be estimated: the use of a penalty information criterion, such as the Akaike or the Bayesian Information criterion, for evaluating the model's fit helps controlling for the risk of overfitting the data. Detailed information on ARIMA and state space models can be found in Bergel-Hayat (2012), Bergel-Hayat and Zukowska (2015a, 2015b), Commandeur et al. (2013) and Dupont & Martensen (2007).

5. An evaluation of the effectiveness of road safety programmes in Australia was performed by Australian researchers who participated in ICTSA, in close collaboration with the Dutch researchers on certain topics.

In the case of Australian research, the main objective was to assess the effectiveness of the major road safety countermeasures. Efforts were made in the following directions: for measuring the effectiveness of police traffic enforcement in different police regions in Victoria (Diamantopoulo & Cameron, 2001; Oppe & Bijleveld, 2003); for assessing the speed camera programme in Victoria (Cameron & Delaney, 2008; Cameron, Newstead, Diamantopoulou, & Oxley, 2003); for estimating the benefits of a vehicle replacement programme in Victoria (Oxley, Cameron, & Newstead, 2003); and for measuring the influence of vehicle types on the total safety of the light passenger vehicle fleet in Australia (Newstead, Delaney, Watson, & Cameron, 2004).

The analysis was explanatory and calls for different methods, ranging from regression to state space methods. In particular, the use of state space methods is discussed in a Ph.D. thesis (Gould, 2005), where models with a range of socio-economic explanatory variables are applied to Victorian accident data. It shows that it is unrealistic to expect models to produce sensible regression relationships unless data are appropriately disaggregated and the model is carefully structured.

Regarding the other analysis, the development of the procedure to measure traffic enforcement effectiveness (the Police Effectiveness Index and the Output Performance Index) was achieved using structural time-series regression techniques. It was concluded that the application of these techniques offers increased power and flexibility in the modelling of crashes compared to traditional multivariate regression methods.

For the assessment of the speed enforcement programmes, the relationship between enforcement and publicity intensity and programme outcomes was analysed. It was shown that both the measures had a statistically significant influence on the reduction of fatal accidents in Victoria.

6. As mentioned above, forecasts can be provided with different sorts of time-series analysis techniques. According to Broughton, and Knowles (2010), it is not clear which method provides the most reliable forecast for 10 or more years in the future.

7. In the case of harmonized data sets, following the same definitions or obtained through the same methodology, the analysis was reliable and identified groups of countries with similar characteristics. Comparative analysis was, however, limited when the data sets used were comparable but not harmonized. This emphasizes the need for more efforts in order to obtain databases of risk factors that really are harmonized.