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®
Results of this study were obtained separately for the composite effects at the camera sites and reference sites analyzed for spillover effects. As best as could be determined, RLC installation was the only change that occurred at the sites when the cameras were installed. In a few cases, the intersection was substantially changed in either the before- or after-period. In such cases, the data after a change in the after-period and before a change in the before-period were excluded from the analysis; therefore, the reported effects most likely result from RLC installation.
Because the intent of the research was to conduct a multijurisdictional study representing different locations across the United States, the aggregate effects over all RLC sites in all jurisdictions was of primary interest. Table 13 shows the combined results for the seven jurisdictions. As seen, there is a significant decrease in right-angle crashes but a significant increase in rear end crashes. Note that "definite injury" crashes here are defined as K, A, and B crashes; they do not contain the "possible injury" crashes captured by KABCO-level "C." Table 14 indicates that, while the magnitude of the effects changes across jurisdictions, the direction of the effects is remarkably consistent.
Note: A negative sign indicates a decrease in crashes.
To investigate possible spillover effects of RLC programs, a separate analysis was performed using the untreated signalized intersection reference sites. For this analysis, the before-and after-periods for these sites in each jurisdiction were demarcated by the year of the first RLC installation at the treatment sites. (Because, by definition, specific treatment dates do not exist for each untreated reference site, this decision was based on the assumption that the public may have perceived that cameras were at noncamera locations from the time of the initial publicity campaign.) Table 15 shows the composite results of this analysis combining data from all of the jurisdictions. As seen, there are indications of a modest spillover effect on right-angle crashes. That this is not mirrored by the increase in rear end crashes that was seen in the treatment group would detract somewhat from the credibility of this result as evidence of a general deterrence effect.
Note: A negative sign indicates a decrease in crashes.
This statistically defendable study found effects that were consistent in direction with those found in many previous studies, although the benefits were somewhat lower than those reported in many sources. This indicates that regression to the mean might have been at play in many of those studies, and it emphasizes the need for controlling those effects in an evaluation of red-light-camera programs, and studies of road safety countermeasures in general.
The opposite direction effects for rear end and right-angle crashes deserves attention from two perspectives. First, the extent to which the increase in rear end crashes negates the benefits for right-angle crashes is unclear at this point. An examination of the changes in crash numbers is insufficient to provide clarity on this issue because of differences in severity levels between right-angle and rear end crashes and in the changes in these crashes following RLC installation. The economic analysis, discussed in the next section, examines the economic costs of the changes based on an aggregation of rear end and right-angle crash costs for various severity levels.
The second perspective of the opposing effects for the two crash types is the implication that RLC systems would be most beneficial at intersections where there are relatively few rear end crashes and many right-angle ones. To provide better guidance on this issue requires an examination of the net economic effect on intersections grouped by the numbers of each crash type. That examination, which seeks generally to isolate all of the factors that would favor (or discourage) the installation of RLC systems, uses the net economic benefit as the outcome variable, also discussed in the next section.
The indications of a spillover effect point to a need for a more definitive study of this issue. That more confidence could not be placed in this aspect of the analysis reflects that this is an observational retrospective study of RLC installations that took place over many years, and in locations where other programs and treatments may have affected crash frequencies at the spillover study sites. A prospective study with an explicit purpose of addressing this issue appears to be required.
This study needed economic cost per crash for the categories of interest. Those categories included "right-angle," "rear end", and "other" at urban and rural signalized intersections. The crash cost needed to be keyed to police crash severity (KABCO) found in the files available for use. In addition, because of limited sample sizes for fatal and severe (A) injury crashes in the after-period for some study intersections, crash costs were needed for combined categories such as K+A severity.
The Pacific Institute for Research and Evaluation (PIRE) developed the cost estimates used in this RLC analysis as part of a larger effort of producing cost estimates for other crash types. Details of the development of the unit crash-cost estimates can be found in a recent paper and in an internal report available from FHWA.(5,6). In summary, by merging previously developed costs per victim keyed on the AIS injury severity scale into U.S. traffic crash data files that scored injuries in both AIS and KABCO scales, PIRE economists could produce estimates for both economic (human capital) costs and comprehensive costs per crash. The comprehensive cost estimates include both economic costs and costs associated with losses in the quality of life. In addition, the analysis produced an estimate of the standard deviation and the 95-percent confidence intervals for each average cost. Following is a list of databases PIRE used:
To meet the needs of this project and future FHWA projects, both comprehensive and human capital cost estimates were developed for six KABCO groupings within 22 selected crash types and two speed limit categories (≤ 72.42 km/h [45 mi/h] and > 80.47 km/h [50 mi/h]). As indicated earlier, the KABCO groupings ranged from detailed estimates for each level of crash severity within each crash type to combined levels of KABCO without regard to crash type. All estimates were stated in Year 2001 dollar costs.
Because this RLC analysis involved placing a value on fatal crashes, comprehensive cost estimates (which include quality-of-life losses) were used as recommended in Council, et al.(3) This specific analysis was focused on right-angle and rear end crashes at signalized intersections in urban areas. Speed limits of ≤ 72.42 km/h (45 mi/h) and > 80.47 km/h (50 mi/h) were used as surrogates for urban and rural here because it was not possible to define an urban/rural variable in the databases PIRE used. Even though a limited number of the study locations had speed limits of 80.47 km/h (50 mi/h), urban unit costs were assigned to all crashes in the analysis. The effect of this approximation probably is small because only 10 of the 132 sites had speed limits of 80.47 km/h (50 mi/h) or more, and all were located in urban areas. Because the initially developed cost estimates for B- and C-level rear end crashes indicates some anomalies in the order (e.g., C-level cost were higher, probably because on-scene police estimates of "minor injury" often ultimately include expensive whiplash injuries), the B- and C-level costs were combined by PIRE into one cost. In initial economic analysis, an attempt was made to use three cost categories within each of the pertinent crash types, K+A, B+C, and no-injury. (It is not feasible to analyze fatal injuries separately in a study such as this because there were limited fatal crashes in any time period. The cost of one fatal crash in any cell could significantly bias the results.) However, because of the low sample sizes of fatal and serious (A-level) crashes in the after-period for some intersections, and the need to use the same cost categories across all intersections in all seven jurisdictions, two crash cost levels were ultimately used in all analyses, injury (K+A+B+C) and non-injury (O). The original estimate developed by PIRE and the combined weighted cost-per-crash estimates used for each crash type are shown in table 16. Also shown in table 16 are the standard deviations for the 2 severity categories used in the analyses.
by severity level and combined weighted estimates used in the economic effects analysis.
Table 17 gives the results for the economic effects including and excluding PDO crashes, estimated from equations 7 to 10 and the associated procedures shown earlier. The latter estimates are included because several jurisdictions considerably underreport PDO collisions. The columns labeled "All crashes" include non-right-angle, nonrear end "other" crashes for which reliable unit costs could not be developed by PIRE because of small sample sizes. It was decided that the same costs as for angle crashes would be used for the "other" category. For completeness, the small changes in these other crashes needed to be accounted for in reporting effects on all crashes, even though the changes may be random and have nothing to do with RLC installation. The results show a positive aggregate economic benefit of more than $14 million over approximately 370 site years, which translates into a crash reduction benefit of approximately $38,000 per site year. The implication from this result is that the lesser severities and generally lower unit costs for rear end injury crashes together ensure that the increase in rear end crash frequency does not negate the decrease in the right-angle crashes targeted by red-light-camera systems.
*Using a combined unit cost for K+A+B+C
As noted earlier, sample size considerations forced the combination of all injury crashes into one category (K+A+B+C). Concern was raised that the distribution of crash severity within this combined category might have changed between the before-and-after periods for either or both crash types. That is, injury-related angle crashes could have become more or less severe between the two periods. If so, the use of one injury-crash cost for both periods would be questionable. Table 18 presents the distributions of both right-angle and rear end injury crashes in each period.
Table 18. Severity-level distributions for right-angle and rear end injury crashes in the before-and-after periods.
As can be seen, while there is no apparent shift in the severity distribution for rear end injury crashes from before to after, the right-angle crashes appear slightly more severe in the after-period (i.e., the percentages of K, A, and B are slightly higher). Further analysis revealed that this shift occurred in only two of the seven jurisdictions. Nevertheless, because the same crash costs essentially were used for before-and after-periods, this means that the cost of the after-period right-angle crashes may be slightly underestimated, even when all jurisdictions are combined. An attempt was made to estimate the potential effect of this shift on the economic savings, even though this could be done only by using anomalous data for individual KABCO categories that we argued against using earlier. With these data, it appears that if the shift were real, the overall cost savings reported in the last row of table 17 could be decreased by approximately $4 million; however, note that there would still be positive economic benefits, even if it is assumed that the unit cost shifts were real and correctly estimated.
Examination of the aggregate economic effect per after-period year for each site indicates substantial variation, much of which could attributed to randomness. It was reasonable to suspect that some of the differences may be to the result of factors that affect RLC effectiveness. The results of the examination of those factors are described next.
Two types of disaggregate analyses were undertaken to identify factors associated with the greatest economic benefits or that might discourage the use of RLCs. The basic outcome measure used is the aggregate economic effects, that is, the combined economic effects on rear end, right-angle, and other crashes of various severities. The economic effect for each crash type and severity was derived from equation 5 as the difference between the expected cost of crashes in the after-period if no RLC were installed and the cost of crashes actually occurring at the treatment sites in the after-period.
The first analysis was a univariate exploration of the results of aggregate economic effects for each intersection, aiming to identify factors that might be associated with the variation in the effects at individual sites. In this, spreadsheets were used to sort the data and results for each site by various columns, and to group by ranges of a variable to explore the relationship between factors and the measured aggregate economic effect per after period site year for a group as a whole for all crash types combined. For example, sorting by ascending order of AADT, the spreadsheet was set up so that aggregate economic effect per after-period site year in a given row applies to all sites with AADT less than or equal to the value in that row; one can then look for trends in the outcome measure. Similarly, by sorting by publicity level, one can obtain the aggregate economic effect per after-period site year separately for the 85 sites with high publicity level and the 47 sites with medium publicity level.
Naturally, some of the conclusions from the univariate exploratory analysis could result from correlation among the various variables found to affect the RLC effect. This could mask the effects or indicate effects that are not real. The results of the exploratory analyses were used to guide a more formal analysis that used multivariate modeling to assess whether the conclusions from the univariate analysis might remain despite the obvious correlations among variables found in that analysis to be associated with economic effects.
In this more formal disaggregate analysis, data for all jurisdictions were combined to develop a model to estimate the value of aggregate economic effect per site year for an individual site using traffic volumes and other site characteristics, such as proportion of rear end or right-angle crashes and signalization features, and RLC implementation features such as publicity level as explanatory variables. The model was a linear one with a normal error distribution. It took the following form:
where α is the calibrated intercept and b1, b2, ... , bn are the estimated effects on Фcost per after-period year of factors x1, x2, x3 ...xn.
Stepwise linear regression was performed with the SAS statistical analysis software package, using the estimates of the Фcost per after-period year as estimates of the dependent variable. It should be pointed out that the absence of a variable in the final model does not necessarily mean that the variable would not affect the safety effect of RLCs because an effect with low statistical significance could result from correlation with other variables, a lack of variation in the data, or a sample that is too small. In addition, it should be emphasized that the generally small size of the aggregate economic effect of RLCs was already strongly indicative of the reality that one is unlikely to detect with significance many factors that affect the safety effect of RLCs.
Data for all treated intersections in all seven jurisdictions were used in this analysis. However, the different jurisdictions had different crash reporting thresholds, which resulted in significantly different numbers and percentages of non-injury crashes across jurisdictions. Because this analysis required that the crash costs for all intersections (and thus all jurisdictions) be calculated on a common basis, non-injury crashes were omitted from this analysis. Because the analysis is aimed at identifying factors of interest, and because these factors can be identified as logically with injury crashes as with total crashes, this was felt to be proper procedure.
The exploratory univariate analysis led to the following general observations on the net economic effects:
Clearly some of these variables that indicate effects in the univariate analysis are correlated, and therefore they may show effects that are not real. For example, left-turn protection is likely related to traffic volume levels; high publicity levels may exist in jurisdictions with the highest traffic volumes. To mitigate this difficulty, the multivariate regression analysis was undertaken to see if the direction of the effect of a given variable remains the same if the effects of other variables are considered simultaneously. This additional analysis confirmed the direction of all of the effects observed except for the penalty variable and the one related to the presence of a left-turn protected phase.
In interpreting these results, consideration should be given to the following important points:
Topics: research, safety, intersection safety, Stop Red Light Running Program
Keywords: research, safety, red light camera, Empirical Bayes, crash evaluation, economic analysis, signalized intersection
TRT Terms: Electronic traffic controls--Evaluation, Photography in traffic engineering, Cameras, Roads--Interchanges and intersections--Safety measures, Traffic safety--United States, Red light running, Cameras, Before and after studies, Economic analysis, Accident analysis, Accident characteristics