measures such as speed limits have a more significant effect on train–vehicle collision severity than on frequency. Although most published work on hot spot identification focuses mainly on developing accident frequency and consequence models separately, few screening studies have proposed frameworks that integrate both elements in a two-dimensional risk approach includ- ing uncertainty in the analysis [examples are Nassar et al. (9) and Saccomanno et al. (10)]. This approach assumes that accident occur- rence at a location is best represented by the product of accident fre- quency and severity. One way to incorporate accident severity in the analysis is to calibrate statistical models that relate accident conse- quences to factors such as location configuration, roadway align- ment, speed limits, and surface conditions (9–11). In this stage, the aim is to identify factors that largely influence the likelihood of fatal or injury outcomes once an accident takes place. For the severity analysis, several statistical model settings have been suggested in the literature, such as the basic logistic regression, multinomial, ordered logit, and mixed logit models [examples are work by Milton et al. (11) and Eluru et al. (12)]. Alternatively, some studies incorporate accident consequences by simply classifying accident counts by severity type (e.g., fatal and injury and other accident types) [exam- ples are work by Miaou and Song (6) and Park and Lord (13)]. In this case, a statistical multivariate model setting considering the different categories is implemented (multivariate analysis). Although this approach accounts for correlation among crash counts, the expected crash consequences (for drivers and passengers) do not vary across locations, and vehicle occupancy levels as an important determinant of overall risk exposure are ignored. Note that vehicle occupancy has been an important part of transportation management systems and is used for evaluating high-occupancy-vehicle lanes or congestion reduction strategies (14). However, vehicle occupancy levels as a determinant of traffic risk exposure have often been ignored in the implementation and evaluation of traffic safety strategies. This paper introduces a new hierarchical Bayesian framework to integrate accident frequency, severity, and vehicle occupancy levels in the hot spot identification process. The primary intention is to illus- trate the potential effect of incorporating accident severity on the result of the hot spot identification process. For this purpose, a group of highway–railway crossings from Canada is used as an application environment. TOTAL RISK–BASED APPROACH In this section, the elements of the proposed Bayesian risk-based methodology are defined, including severity score, accident consequence model, and hot spot identification criteria. How to Incorporate Accident Severity and Vehicle Occupancy into the Hot Spot Identification Process? Luis F. Miranda-Moreno, Liping Fu, Satish Ukkusuri, and Dominique Lord 53 This paper introduces a Bayesian accident risk analysis framework that integrates accident frequency and its expected consequences into the hot spot identification process. The Bayesian framework allows the intro- duction of uncertainty not only in the accident frequency and severity model parameters but also in key variables such as vehicle occupancy levels and severity weighting factors. For modeling and estimating the severity levels of each individual involved in an accident, a Bayesian multinomial model is proposed. For modeling accident frequency, hier- archical Poisson models are used. How the framework can be imple- mented to compute alternative relative and absolute measures of total risk for hot spot identification is described. To illustrate the proposed approach, a group of highway–railway crossings from Canada is used as an application environment. Because of the deficiencies of accident risk estimates based on raw data, the traffic safety community is interested in the development and application of the risk model–based approach, which makes use of statistical methods based on probability theory. The approach con- sists of a systematic analysis of the input crash data to develop acci- dent frequency and consequence models from which ranking criteria are built (1–3). Once statistical models have been developed from the input data, several Bayesian ranking methods or criteria proposed in the literature can be applied to identify a list of hot spots (1–8). These criteria include the posterior expectation of accident frequency, the potential of accident reduction, and the posterior expectation of ranks. These measures usually are based on the assumption that the safety status of a site can be reflected by accident frequency, and severity is usually not incorporated in the analysis or is assumed to be fixed over locations (observed and unobserved severity heterogeneities are ignored across sites). In many applications, however, accident frequency may not completely reveal the total risk level of a site or capture the safety benefits that some safety countermeasures could introduce. For example, in highway–railway networks, some safety L. F. Miranda-Moreno, Department of Civil Engineering and Applied Mechanics, McGill University, Macdonald Engineering Building, 817 Sherbrooke Street West, Montreal, Quebec H3A 2K6, Canada. L. Fu, Department of Civil and Environmen- tal Engineering, University of Waterloo, 200 University Avenue West, Waterloo, Ontario N2L 3G1, Canada. S. Ukkusuri, Department of Civil and Environmental Engineering, Rensselaer Polytechnic Institute, 4032 Jonsson Engineering Center, Troy, NY 12180. D. Lord, Zachry Department of Civil Engineering, Texas A&M University, 3136 TAMU, 317 Gilchrist Building, College Station, TX 77843-3136. Corresponding author: L. F. Miranda-Moreno, luis.miranda-moreno@mcgill.ca. Transportation Research Record: Journal of the Transportation Research Board, No. 2102, Transportation Research Board of the National Academies, Washington, D.C., 2009, pp. 53–60. DOI: 10.3141/2102-07