Several different factors contribute to injury severity in traffic accidents, such as driver characteristics, highway characteristics, vehicle characteristics, accidents characteristics, and atmospheric factors. This paper shows the possibility of using Bayesian Networks (BNs) to classify traffic accidents according to their injury severity. BNs are capable of making predictions without the need for pre assumptions and are used to make graphic representations of complex systems with interrelated components. This paper presents an analysis of 1536 accidents on rural highways in Spain, where 18 variables representing the aforementioned contributing factors were used to build 3 different BNs that classified the severity of accidents into slightly injured and killed or severely injured. The variables that best identify the factors that are associated with a killed or seriously injured accident (accident type, driver age, lighting and number of injuries) were identified by inference.Research highlights▶ Bayesian Networks are usefully applied in the domain of traffic accident modeling. ▶ BNs are used for classifying traffic accidents according to their injury severity. ▶ BNs inference identifies variables associated with KSI (killed or seriously injured). ▶ The key variables for KSI were accident type, age, lighting and number of injuries.
The number of traffic accidents and their effects, mainly human fatalities and injuries, justify the importance of analyzing the factors that contribute to their occurrence. Identifying the factors that significantly influence the injury severity of traffic accidents was the main objective of many previous studies. Factors affecting injury severity of a traffic accident are usually caused by one or more of the following factors: driver characteristics, highway characteristics, vehicle characteristics, accidents characteristics and atmospheric factors (Kopelias et al., 2007 and Chang and Wang, 2006).
Regression analysis has been widely used to determine the contributing factors that cause a specific injury severity. The most commonly used regression models in traffic injury analysis are the logistic regression model and the ordered Probit model (Al-Ghamdi, 2002, Milton et al., 2008, Bédard et al., 2002, Yau et al., 2006, Yamamoto and Shankar, 2004 and Kockelman and Kweon, 2002). However, most of the regression models that are used to model traffic injury severity have their own model assumptions and pre-defined underlying relationships between dependent and independent variables (i.e. linear relations between the variables) (Chang and Wang, 2006). If these assumptions are violated, the model could lead to erroneous estimations of the likelihood of severe injury.
Gregoriades (2007) highlighted the interest of using Bayesian Networks (BNs) to model traffic accidents and discussed the need to not consider traffic accidents as a deterministic assessment problem. Instead, researchers should model the uncertainties involved in the factors that can lead to road accidents. He listed a number of candidate approaches for modeling uncertainty, such as, Bayesian probability.
BNs make it easy to describe accidents that involve many interdependent variables. The relationship and structure of the variables can be studied and trained from accident data. They do not need to know any pre-defined relationships between dependent and independent variables.
The three main advantages of BNs are bi-directional induction, incorporation of missing variables and probabilistic inference. By using BNs, it is relatively easy to discover the underlying patterns of data, to investigate the relationships between variables and to make predictions using these relationships. Incident data used in a study by Ozbay and Noyan (2006) were collected from incident clearance survey forms to understand incident clearance characteristics and then used to develop incident duration prediction models. The researchers modeled the incidents’ clearance durations using BNs and were able to represent the stochastic nature of incidents.
Using BNs to analyze traffic accident injury severity is scarce. A two car accident injury severity model was constructed using BNs (Simoncic, 2004). A BN was built using several variables, and the Most Probable Explanation (MPE) was calculated for the most probable configuration of values for all the variables in the BN, in order to serve as an indication of the quality of the estimated BN. The results pointed out that BNs could be applied in road accident modeling, and some improvements, such as using more variables and larger datasets, were recommended. Although this study highlighted the possibility of using BNs to model traffic accidents, the results were based on building only one possible network, without measuring the performance of the Bayesian classifier.
The scope of this paper is to validate the possibility of using BNs to classify traffic accidents according to their injury severity, and to find out the best BN classification performance along with the best graphical representation, in order to be capable of identifying the relevant variables that affect the injury severity of traffic accidents.
This paper is organized as follows. Section 2 presents the data used and briefly reviews the concept of BNs and Bayesian learning. The methods used for preprocessing and evaluating the data are also presented; finally a brief description of inference is presented. In Section 3, the results and their discussion are presented. In Section 4, summary and conclusions are given.
This paper uses BNs to analyze traffic accident data in order to validate the ability of this data-mining technique to classify traffic accidents according to their injury severity, and to identify the significant factors that are associated with KSI in traffic accidents.
Traffic accident data was obtained from the DGT for a period of three years (2003–2005) for Granada (Spain). Three BNs were built using three different score metrics: BDe, MDL and AIC.
Several indicators have been used in order to evaluate the performance of the built BNs: accuracy, sensitivity, specificity, HMSS, ROC Area, MPE and graph complexity (or number of arcs). The results obtained for these indicators do not vary significantly between the different score metrics used and they are within the range of previous studies (Abdel Wahab and Abdel-Aty, 2001 and Simoncic, 2004). So, it could be concluded that BNs might be a useful tool for classifying traffic accidents according to their injury severity.
Inference was used to identify the values of the variables that are associated with KSI in traffic accidents on Spanish rural highways. Based on the results, it would be possible to identify the type of accident that would most probably be classified as KSI on Spanish rural highways. It would be a head-on or rollover traffic accident in a roadway without lighting with only one injury within the age of 18 and 25 years. These factors (head-on or rollover, unlit roadway, only one injury and within the age of 18 and 25 years) do not have to exist all at once in order to have a KSI accident. Any of these or a combination of them might increase the probability of a KSI accident. In general, these results are consistent with the literature (Tavris et al., 2001, Kockelman and Kweon, 2002, Abdel-Aty, 2003, Helai et al., 2008, Gray et al., 2008 and Scheetz et al., 2009). However, this finding may vary for other countries and datasets.
BNs, which have proved their effectiveness in different research areas, could be usefully applied in the domain of traffic accident modeling. Their effectiveness has been found to be similar to other data-mining techniques used to model severity in traffic accidents. Compared with other well-known statistical methods, the main advantage of the BNs seems to be their complex approach where system variables are interdependent and where no dependent and independent variables are needed (Simoncic, 2004).
