Featuring developments in Federal highway policies, programs, and research and technology.
|This magazine is an archived publication and may contain dated technical, contact, and link information.|
|Federal Highway Administration > Publications > Public Roads > Vol. 59· No. 2 > Interactive Highway Safety Design Model: Accident Predictive Module|
Interactive Highway Safety Design Model: Accident Predictive Module
by Harry Lum and Jerry A. Reagan
The Summer 1994 issue of Public Roads introduced the Interactive Highway Safety Design Model (IHSDM). (1) (See figure 1.) This tool, which is being developed by the Federal Highway Administration (FHWA) at the Turner-Fairbank Highway Research Center (TFHRC), will provide to highway design engineers safety information on the relationships between geometrics and accidents in a usable format and will guide the designer in evaluating the safety of alternative alignments. This article discusses the accident predictive module of IHSDM.
There have been a substantial numbers of efforts in the past to relate accidents with geometric features. These studies assumed there was a definable relationship between geometric design elements and traffic accidents. An excellent summary of the results of many of the accident studies has recently been published in a six-volume report. (2) This report is a classic example of three major reasons why many of the safety relationships have not been included in design manuals.
First, the studies focused on specific issues and did not examine the roadway as a whole. Each of the six volumes deals with a specific aspects of geometric design. Consequently, the information is not presented as a comprehensive design procedure but as a series of safety predictive relationships that address specific issues.
Second, the safety predictive models relating specific geometric features to accidents are not in a format used by designers. The results of safety research studies need to be expressed in a format that can be directly incorporated into the American Association of State Highway And Transportation Officials (AASHTO) policy. (3) The new guidelines must be compared with current AASHTO safety policy and shown to be superior.
Thirdly, the predictive models used R2, the coefficient of determination, to judge the models' accuracy. Predictive equations with R2 values in the 0.25 to 0.35 range indicate that the variables selected in the model explain 25 to 35 percent of the predicted values and that 65 to 75 percent of the phenomenon remains unexplained. The low R2values associated with these models have troubled design engineers for years. Most design engineers have a fairly limited background in statistics and distrust low R2 values.
As shown in figure 1, the IHSDM architecture connects the accident predictive module to the data stored in the computer-aided design (CAD) package. Eventually, IHSDM will allow the other modules shown in figure 1, especially the driver module and the traffic module, to communicate with the accident predictive module. This flexibility provides an architecture that overcomes the shortcomings of earlier studies. From a research standpoint, it provides a framework that can be used to both design safety studies and to allow the results to be evaluated as part of a total highway project. From a designer standpoint, by simply running the accident predictive module, an average accident rate for any given alternative can be calculated without having to enter any additional data.
The development of the accident predictive module will be an evolutionary process in three phases.
Phase one is the short-term goal and deals only with two-lane, rural roads. This phase will use the results presented in an FHWA study on two-lane, rural roads. (4,5) The regression safety relationships presented in the FHWA study will be computerized and adapted to use the CAD x,y,z, format. This phase has just gotten under way in the Geometric Design Laboratory at TFHRC. Phase 1 will provide insight into what sort of problems are involved in using the CAD format and safety relationships. It is anticipated that this phase will be completed in 1996.
Phase 2 is also under way and is a medium-term goal. It is a continuation of phase 1 and will address both roadway and roadside accidents and will continue to develop the interface between the accident predictive module and the CAD package. Phase 2 will be the development of the accident predictive module described in a recent FHWA study. (6) This module is shown in figure 2 and consists of four submodels -- roadways (nonintersections), intersections, interchange ramps and roadside.
Three of the submodels -- roadways, intersections and interchange ramps -- will be developed by FHWA through staff and contract studies. Experimental designs for these three submodels are being developed under two FHWA contracts awarded in 1992. Currently, the experimental plan for the intersection accident predictive submodel is under way. (See sidebar.) The roadside submodel is being revised and updated under National Cooperative Highway Research Program (NCHRP) Project 22-9 (which like the computer program, ROADSIDE) will be an encroachment based procedure involving a chain of conditional probabilities. (7) One FHWA contractor is currently trying to develop an experimental design that can be used to validate the encroachment values used in NCHRP 22-9. The overall accident rate of the highway will be the sum of the four submodels.
Phase three is a long-term goal. To improve our ability to predict accidents, new techniques need to be developed. For example, future accident prediction modules could be based on simulations involving the geometric design elements that are stored in the CAD package, the traffic module, and the driver module. This appears to be the only way to deal with the third issue discussed earlier - low R2 values. Accidents are relatively rare events in terms of the number of vehicles and drivers on the road. The interactions that can occur from an accident are very complicated. Some are related to the drivers; some to the vehicle; some to the roadway; and all are related to the natural environment (weather, ambient light conditions) over which the design engineers have little control. This interaction complicates the task of sorting out the safety effect attributable to a specific geometric feature. However, there may be an explanation for the low R2 values. Figure 3 presents some findings from a study that compares causes of accidents in the United States and Great Britain. (8)
This study notes that only 3 percent of accidents are due solely to the roadway environment, 57 percent solely to drivers, 2 percent solely to vehicles, 27 percent to the interaction between road environment and drivers, and 3 percent to the interaction of the environment-driver vehicle. Taken at face value, this suggest that road-related elements are associated with 34 percent of crash causation. Consequently, a perfect model would attribute about 34 percent of R2 to the roadway road variables, including the driver and the vehicle.
Recent work in this area looks at an alternative methods of evaluating the effects of the driver and the traffic. (9) One obvious question is: Why not go directly to phase 3? There are three reasons. While there are traffic models that could (and will be) adapted to IHSDM, there is no appropriate driver module. The analysis and simulation techniques have not been developed. Finally, the accident relations from phase 2 may be needed to validate the simulation models.
A new approach is needed. A methodology is needed to establish the relationship between geometric design elements, accidents, and driver performance.
Previous safety studies that have attempted to related safety and geometrics design elements have not been very satisfactory. Past studies using statistical analysis of existing data bases have produced little results helpful to the design engineer. Even special studies that have collected all new data over extended periods of time have not been successful in clearly defining the contribution of geometrics to accidents. The tentative conclusion drawn here is that the design of geometric elements is an engineering and drivers' behavior problem; neither can be treated independently of the other. Methodologies must be developed that account for the role of driver behavior in accidents.
Status of the Accident Predictive Module: Intersection Submodel
The first of the FHWA submodels to be developed is the intersection submodel. The current effort deals only with the development of the experimental design and its review.
The experimental design of this submodel began in late 1993 and should be concluded in 1994. The experimental design is trying to answer such questions as analysis techniques, submodel form, data needs, and costs associated with the development of the submodel. The experimental design began with the evaluation of existing data bases that had a history of accident experience related to geometrics. California's data base was determined to be the best available although it was far from complete. For example, it does not include data on traffic volumes for various turning movements at intersections and sight distances.
After editing and deleting low-volume intersections -- those used by fewer than 400 vehicles per day -- the remaining intersections were classified as either four-legged, rural with stop control; four-legged, urban with stop control; three-legged, rural with stop control; three legged, urban with stop control; and four-legged, urban with signalized control. Stop control in all cases was on the minor approach. Accident was defined as one that occurred within the curbline limits of the intersections and within 75 meters of the intersections. The response variables were: (1) total multiple-vehicle accidents and (2) fatal and injury multiple-vehicle accidents. Between 10 to 12 variables from a list of 20 variables were chosen on the basis of engineering judgment as independent variables.
The five intersection types were subdivided into three groups:
The statistical R2 was used as the criterion of goodness of fit -- the higher the proportion of variances explained by the input independent variables, the better the fit. Table 1 summarizes the results of the different models. With the exception of the signalized, urban intersection, the logistic fared much better than the Poisson. The logistic model as with the Poisson model, however, gave some anomalous prediction -- e.g., lighting increases accident experience.
In an effort to improve the accuracy of the predictive models and remove some of the anomalous values, additional variables have been collected that were not available in the original data base. Specifically, several variables were identified for the urban, signalized, four-legged intersection model - turning movement volumes during the morning and evening peak hours, intersection sight distance, and related geometric data. Data were collected for 200 intersections in the areas of the Sacramento Valley, San Francisco Bay, and Southern California. Currently, these are being subjected to regression analysis to see if adding these variables improves the preliminary models and removes anomalous predictions. Once this analysis is completed, the experimental plan will be finalized. The experimental plan will identify the types of data that must be collected, the amount of data to be collected, the costs associated with that data, and the analytical techniques that will be used to develop the predictive model.
An expert panel will be convened at the TFHRC in early 1995 to review the experimental plan. One of the key questions the panel must address: Should the development of the intersection submodel be continued? It is anticipated that the experimental plan will provide predictive submodels whose R2 values are not significantly higher than past studies. However, the submodel represents the current state of the art and can be developed in a reasonable time period. One of the major reasons for continuing is the need for a transition model from current technology to more rigorous analytical methods. An interesting alternative approach that uses a traffic conflicts, computer simulation model has recently been reported. (9) The model provides a means of studying traffic conflicts and the effect of driver and traffic parameters on the occurrence of conflicts. Missing is the link between traffic conflicts and accidents.
Table 1 -- Results (R2) of Logistic and Poisson Regression Models of Intersection Accidents for California, 1990-92
TMV = total multiple vehicle accidents
FIMV = fatal/injury multiple vehicle accidents
*Regression based on log-normal.
(1) Jerry A. Reagan. "The Interactive Highway Safety Design Model: Designing for Safety by Analyzing Road Geometrics," Public Roads, Vol. 58, No. 1, Federal Highway Administration, Washington, D.C., Summer 1994.
(2) Safety Effectiveness of Highway Design Features, Vol. I-VI. Publication Nos. FHWA-RD-91-044 to -049, Federal Highway Administration, Washington, D.C., 1992.
(3) A Policy on Geometric Design of Highways and Streets, American Association of State Highway and Transportation Officials, Washington, D.C., 1990.
(4) Design Risk Analysis, Vol. I, Publication No. FHWA-FLP-010, Federal Highway Administration, Washington, D.C., April 1991.
(5) Design Risk Analysis, Vol. II, Publication No. FHWA-FLP-011, Federal Highway Administration, Washington, D.C., April 1991.
(6) Conceptual Plan for an Interactive Highway Safety Design Model, Publication No. FHWA-RD-93-122, Federal Highway Administration, Washington D.C., February 1994.
(7) K.K. Mak and D.L. Sicking. Improved Procedures for Cost-Effectiveness Analysis of Roadside Safety Features, National Cooperative Highway Research Program Project 22-9, publication pending.
(8)K. Rumar. "The Role of Perceptual and Cognitive Filters in Observed Behavior," Human Behavior in Traffic Safety, eds. L. Evans and R. Schwing, Plenum Press, 1985.
(9)T. Sayed, G. Brown, and F. Navin. "Simulation of Traffic Conflicts at Unsignalized Intersections With TSC-Sim," Accident Analysis and Prevention, Vol. 26, No. 5, 1994, pp. 593-607.
Harry Lum retired on Oct. 1, 1994, after 25 years with FHWA. Most recently, he served as a mathematical statistician in the Design Concepts Research Division, Office of Safety and Traffic Operations Research and Development, at TFHRC in McLean, Va.
Jerry A. Reagan is chief of the Design Concepts Research Division. Prior to this assignment, he served as chief of the Safety Traffic Implementation Division. He has had a variety of experiences with FHWA, beginning in 1967 as a materials engineer. Later he was assigned to Region 15 as a soils and foundation engineer. In 1973, he transferred to the Office of Environmental Policy at FHWA headquarters, where he worked for 10 years. Then he moved to TFHRC as the state programs officer of the National Highway Institute (NHI), where he was responsible for the NHI short-course program. He has a bachelor's degree and a master's degree in civil engineering from the University of Tennessee. He is a registered professional engineer in Tennessee and Virginia.
Page Owner: Office of Corporate Research, Technology, and Innovation Management
Scheduled Update: Archive - No Update
Technical Issues: TFHRC.WebMaster@dot.gov