Developing a Composite Index for Assessing Safety Performance of Drivers Based on Data Envelopment Analysis

Abstract

Transportation safety issues are still severe, despite many efforts that have been made to alleviate the negative impacts they pose to society. In most cases, human factors play a crucial role in the occurrence of traffic crashes. When splitting the number of road fatalities by road user type, drivers represent the largest share. As a result, to improve road safety in the future, it is essential to better understand the behavior of different drivers. Objective driving performance measurements are valuable to identify the safety status of individual car drivers. However, a uniform methodological framework has not yet been successfully established to address these issues. One of the focus areas could be drivers over the age of 65, as the proportion of elderly drivers is increasing considerably. Diminished sensory, motor, and cognitive abilities and increased fragility pose a significant challenge to the traffic safety performance of this user group. Maintaining the balance between mobility and safety of this cohort has been of recent discussion. Nonetheless, there is still no universal method to identify elderlies with diminished driving abilities in the early stages before jeopardizing their own and other people's life. Besides, speeding, driving under the influence of alcohol, and the non-use of protective systems are the main contributing factors to most traffic-related casualties. Thus, this dissertation aims to evaluate drivers' performance, including middle-aged and elderly drivers, and road user behavior in different European countries with scientifically sound methodologies and data-driven approaches. The thesis objectives are outlined as follows: 1. To objectively evaluate drivers' performance, at the individual level, based on the driven speed, acceleration, and lateral position. 2. To assess the driving ability (i.e., fitness-to-drive) of elderly drivers based on the results of a set of functional ability tests. 3. To benchmark countries in terms of the safety performance of road users concerning drink-driving, speeding, and nonuse of protective systems. In order to achieve the abovementioned objectives, several research studies were conducted and provided in different chapters of this dissertation. Chapter 1 is initiated by presenting the background and theoretical framework of the research topics. After a general introduction to the road safety issue, the method of aggregating individual indicators into an index is discussed, and the DEA approach is introduced. In particular, the DEA is explained with a simple numerical example in which the concept of the efficient frontier, production possibility set, efficient and inefficient DMU, and projection points are illustrated utilizing insightful graphs. Then the mathematical form of DEA and some of its extensions, which have been developed and applied in this dissertation, are described. The following five core chapters, as the primary studies in this dissertation, are categorized into three different parts based on their context: Part 1. Driving Performance Index Part 2. Older Driver Ability Index Part 3. Road User Behavior Index The first part deals with evaluating driving performance in a curve-taking scenario. This is performed by executing two case studies (i.e., Chapters 2 & 3). A new framework for constructing a driving performance index is proposed in Chapter 2 by applying the DEA method to a driving simulator data set. The Driving Performance Index evaluates the relative performance of individual drivers based on the three most common parameters used to describe and analyze a driver's behavior, i.e., speed, acceleration, and lateral position. Then in Chapter 3, the proposed framework is integrated with the Window Analysis approach. It is a method for examining the variations in DMUs’ (i.e., the driver in our case) performance over time. In each period, the driver is treated as if they are a different driver. Consequently, a driver's performance in each period is evaluated concerning its performance at other points in time, along with the performance of other drivers. The aim is to consider the multidimensionality of driving performance and its changes over time. Apart from identifying the bestperforming and underperforming drivers, all drivers are ranked based on their calculated index scores. Their relative performance is achieved for speed, acceleration, and lateral position. The driving performance index can select candidates for driving jobs, identify high-risk drivers, improve the rating process, and reward low-risk drivers. It can also be a good tool for insurance companies to use different pricing schemes for other drivers based on their performance. The second part focuses on the fitness-to-drive assessment of older drivers (i.e., Chapters 4 & 5). In Chapter 4, the driving ability of elderly drivers is assessed based on specific measures of functional and driving abilities using DEA and MultiCriteria Decision Analysis (MCDA) methods. In this experiment, participants aged 70 years and above completed tests of an assessment battery of psychological and physical aspects and knowledge of road signs. Additionally, they took a driving simulator test in which specific driving situations known to cause difficulties for elderly drivers were included. The overall ability of each elderly driver is evaluated based on all the above information combined in an index using the Multi-Layer DEA-based cross-index (MLDEA-based cross-index) approach. The model output distinguishes the best-performers from those underperforming drivers. It helps guide the development of training interventions tailored to each individual by explicitly targeting those (mainly) impaired functions. Next, to assess the robustness of the results obtained with the DEA model, the analysis was also performed with one of the MCDA methods, Preference Ranking Organization Method for Enrichment Evaluations (PROMETHEE II). It is done to enhance the study of the problem by using the stability intervals (i.e., outputs of the sensitivity analysis) of PROMETHEE II as a weight restriction in the DEA model. Furthermore, to give the elderly drivers more insight into characterizing their driving abilities, the possibilities and challenges of considering the complimentary use of the MLDEA cross-index model and PROMETHEE II as an analyzing tool in evaluating the overall performance of older drivers were investigated. Spearman's correlation coefficient confirmed the replicability of the results with the PROMETHEE II model and the robustness of the final solutions. Finally, we were able to study the participants' profiles in more detail using the GAIA plane and spider web representations. From a practical point of view, it enables road safety stakeholders and fitness-to-drive evaluators to track elderly’s driving behavior and identify weak areas of their performance, and then develop strategies and remedial actions to eliminate some of those weaknesses. Chapter 5 presents the results of a study in which elderly participants were evaluated on their visual, physical, and cognitive abilities. Furthermore, the same participants completed an on-road driving assessment similar to the Belgian fitness-to-drive evaluation procedure. The dataset contains 19 variables that describe the functional ability of participants and one variable to indicate whether they are fit to drive or not based on their performance in the on-road test. Given the dataset, i.e., 19 independent variables as predictors and the latter categorical variable as the label, the aim was to develop a cost-effective and easy-toadminister prediction model to predict the fitness-to-drive of elderly drivers based only upon the result of a set of functional ability tests. The idea was to use advanced data mining and machine learning techniques rather than traditional statistical approaches. Various supervised binary classification models were trained and tested using different classes of machine learning algorithms. Of all classifiers, the one fitted through the gradient boosting model (GBM) appeared to be the best model due to its highest predictive performance and the largest area under the “receiver operating characteristic” curve of 0.92 and accuracy of 90%. This classifier correctly specified 95% of the fit-to-drive participants as fit (i.e., true positive rate or sensitivity=0.95) and 80% of the unfit-to-drive participants as unfit (i.e., true negative rate or specificity=0.80). This study revealed that the following factors contributed most to the classification of older drivers in the fit vs. unfit classes: helpful field of view measuring the processing speed, functional reach test, audio verbal learning, knowledge of road signs, and visual acuity. In light of this, all these elements must be tested during license renewal, not just the eyesight as it is usually done. The third part (Chapter 6) evaluates road user behavior in a set of selected European countries. The aim is to assess the safety performance of road users regarding the following driving safety-related issues: drink-driving, speeding, and nonuse of protective systems (i.e., seat belts) in a country and draw a meaningful comparison of the road safety situation among other countries. It helps to learn from the best-in-class and speed up safety improvements. In this regard, a new DEA index methodology is proposed in which optimal common weights are determined. At the same time, the model considers the layered hierarchy of the indicators. The suggested approach provides a fair and identical basis for evaluating and comparing road user behavior and reduces the required computational burden for solving the formulated model. The final chapter, Chapter 7, summarizes the main findings of the five core chapters of this dissertation, points out its limitations, and proposes suggestions for future studies in the field of driving evaluation.