U.S. Department of Transportation
Federal Highway Administration
1200 New Jersey Avenue, SE
Washington, DC 20590
Federal Highway Administration Research and Technology
Coordinating, Developing, and Delivering Highway Transportation Innovations
|This report is an archived publication and may contain dated technical, contact, and link information|
Publication Number: FHWA-HRT-05-048
Date: April 2005
Safety Evaluation of Red-Light Cameras
PDF Version (621 KB)
PDF files can be viewed with the Acrobat® Reader®
Before the actual data analyses, preliminary efforts involving file merging and data quality checks were completed.
In most State DOT crash and inventory files, as is the case with the State data in FHWA's Highway Safety Information System (HSIS), crashes can be computer-linked to inventory and traffic volume data for roadway segments and intersections using location reference variables such as route or milepost on each record. This was possible in Howard and Montgomery Counties, where project staff were provided an electronic file of the milepost book used by police officers in the field. The treatment and reference intersections were identified and matched to crashes based on this milepost information.
This was not the case for the intersections analyzed in other jurisdictions. In Charlotte, NC, the 1997 and later data included intersection control numbers for all intersections and crashes that could be used for file linkage. There were also control numbers in the pre-1997 data, but they differed from the later data. The Charlotte staff provided us with conversions between the new and old systems.
For all data for the three California jurisdictions and Baltimore, MD, no such location system existed, and the crashes had to be manually linked to pertinent intersections based on the names of the crossing streets. The crashes were sorted by street names and an analyst matched the crash-report streets with the street names from the treated and comparison intersection file. All combinations of crash-report street names were checked to pick up possible misspellings by the investigating officer or coder.
The project team was able to conduct a limited verification of both the completeness of the State CODES data files and the manual linkage procedures using El Cajon, CA, data. The local traffic engineer sent the project team crash summary reports for one treatment, one signalized reference, and one unsignalized control intersection. These summary reports contain a listing of all cases that have been coded to an intersection by city staff, using their own coding scheme. The comparison of these crashes to those identified and linked by the project team indicated that use of the State data resulted in minor differences with the local crash summaries. It was thus concluded that the State data is of sufficient quality.
As indicated earlier, the basic analyses were to be focused on target crashes, those red-light-running crashes that could be affected by the RLC treatment. The analysis would also examine other intersection crashes to confirm that unanticipated effects were not present. Because there is no "red-light-running" crash category on most police crash forms, these target crashes must be defined based on variables on the form. Definitions could range from only crashes in which a citation for a traffic signal violation was given to all right-angle crashes and rear end crashes at or near the intersection. One could choose to include rear end crashes that were noted by the officer as "intersection-related" (where this variable was present), or to include all rear end crashes approaching the intersection within a specified distance of the intersection. Depending on the distance (X) chosen, the assumption would be that the RLC would affect behavior of the lead vehicle or vehicles, which could result in rear end crashes X distance back in the approaching queue of vehicles. One could also choose to include left-turn opposite-approach crashes because some of these would be red-light-running crashes if a protected signal phase existed. To further complicate matters, the different jurisdictions use slightly different definitions of right-angle crashes on the report form.
Based on definitions used in previous studies, available data variables in the current files, and project team discussions, the following general decisions were made by the project team:
As could be expected, available crash variables and codes differed between cities, making it impossible to have totally consistent definitions across all seven jurisdictions. For example, only the three Maryland jurisdictions had an "intersection-related" code that can be used to further screen rear end crashes occurring within 45.72 m (150 ft) of the intersection. Thus, all rear end crashes within 45.72 m (150 ft) were used in Charlotte, NC, and all three California databases.
In addition, we encountered significant problems with the distance-from-intersection data in Baltimore, MD. Approximately 10 percent to 15 percent of the data appear to have questionable distances such as distances of 0.03 m (0.1 ft) and 0.30 m (1.0 ft) from the intersection. The project team attempted to verify these distances by obtaining hardcopies, but found that the accident case numbers in the computerized CODES data were not the same as the Baltimore Police Department case numbers, and only Baltimore has hardcopies of the reports. Thus, in the Baltimore analyses, two sets of data were used, a first set containing only rear end crashes within 45.72 m (150 ft) where the distance data were believed to be accurate, and a fuller set that also included crashes coded as within 45.72 m (150 ft), where the distance measurements were questionable. The analyses of these two sets of data revealed no significant differences; therefore the full set including the questionable distances was used for the final analysis.
The final set of criteria for each RLC-related crash type for each jurisdiction is listed in table 11.
As indicated earlier, the study required the development of safety performance functions (SPFs) for signalized and stop-controlled intersections. A reference group of untreated signalized intersections was used to develop SPFs to account for traffic volume changes and regression to the mean using the empirical Bayes procedure. The unsignalized intersection SPFs were used to account in that procedure for time trends in crash counts unrelated to the RLC installation. Therefore, it was necessary to first ensure that the comparison group used to calibrate the SPFs was suitable for this purpose, that is, that it had similar crash trends to the treatment group over the years before RLC installation. To this end a comparability test as outlined in Hauer was performed.(4) This test confirmed the suitability of the comparison group.
To build the strongest possible SPFs, reference group data (i.e., data from the untreated signalized intersections) were combined for sets of jurisdictions, considering proximity and similarity in crash reporting practices. To this end, the three California cities of El Cajon, San Diego, and San Francisco were combined. Not only are these three cities in proximity, but they also do not have full reporting of PDO crashes, and the crash data all came from the State database maintained by the CHP. Howard and Montgomery Counties, MD, reference group data were combined because of their proximity and similarity in reporting practices. Baltimore, MD, and Charlotte, NC, were combined because of their high reporting of non-injury crashes. In each case where jurisdictions were combined, a jurisdiction-specific multiplier was calibrated and applied to account for any remaining differences in crash reporting.
Development of the SPFs involved determining which explanatory variables should be used, whether and how variables should be grouped, and how variables should enter into the model, in other words, the best model form. Generalized linear modeling was used to estimate model coefficients using the software package GENSTAT and assuming a negative binomial error distribution, all consistent with the common recent research practice in developing these models.(29)
In specifying a negative binomial error structure, the dispersion parameter, k, which relates the mean and variance of the regression estimate, is iteratively estimated from the model and the data. The value of k is such that the smaller its value, the better a model is for a given set of data.
For specific crash types at signalized intersections, a multiplier is applied to the model that is equal to the proportion of total crashes that each crash type makes up. A value of k was calculated for each crash type using a maximum likelihood process, as explained earlier. Similarly, although data for groups of jurisdictions were combined for SPF calibration, separate multipliers and k values were calculated for each jurisdiction.
The inclusion of variables such as number of lanes rarely significantly affected the fit. This is not surprising because, as previous research has shown, much of the variation in crash experience is explained by the volume of traffic entering an intersection. The results of the SPF calibration for the signalized reference group are presented in table 12. The model forms used are tried and tested and, because of the limited datasets available, options on model forms and variables to include were so limited that a trial and error modeling approach, using published models as a guide, was realistic. In addition, fine tuning the model is not critical in EB analysis, especially because by weighting the observed count, one is accounting for omitted variables that may affect crash frequency.
F1 = entering AADT on major road, F2 = entering AADT on minor road; minllane = number of left-turn lanes on the minor road;
majllane = number of left-turn lanes on the major road; minrlane = number of right-turn lanes on the minor road; (s.e.) = standard error of the estimate;
k is a calibrated parameter relating the mean and variance used in the empirical Bayes estimation procedure