Procedures for the identification of black spots, or hazardous locations, are attempts to select some sites out of many to improve safety. These are sites with promise. The historical and conceptual development of such procedures is reviewed. On the basis of this review, an attempt is made to create some order in the thinking. Suggestions are made to improve identification. It is noted that the status of the stage that follows identification, the stage of site safety diagnosis and remediation, is underdeveloped. This section gives a historical sketch of the main ideas and concep- tual developments in what is usually called the identification of black spots, or hazardous locations. To make this review easier, the following notation will be used: X = observed accident count for a road section and period, m = expected accident count (E{X}) for the road section and period, E{m} = mean of m’s for similar road sections, D = length of road section, Q = number of vehicles passing road section during period to which X pertains, R = observed accident rate (e.g., accidents/vehicle- kilometer) [note that R = X/(DQ)], R EB = accident rate estimated by empirical Bayes method, R  = average value of R for similar road sections, UCL X = upper control limit for observed accident counts (X), UCL R = upper control limit for observed accident rate (R), t = number of years of accident data to be used, and , = parameters. Norden et al. (1) suggest using methods of industrial statistical quality control for highway safety. What is being monitored is the observed accident rate R. They assume that if some road section served D  Q vehicle miles in a certain time period, it would be expected to have R ¯¯ DQ accidents in that period. One can now find an accident count X, called the upper control limit (UCL X ), such that the probability X  UCL X is less than 0.5 percent. Equivalently, UCL R = UCL X /(DQ) is the upper limit on the observed accident rate R. Using an approximation of the Poisson distribution and 0.5 percent probability, they suggest that UCL R = R ¯¯ + 2.576[R ¯¯ /(DQ)] + 0.829/(DQ) + (DQ)/2. Similar approaches were used by Rudy (2) and Morin (3). Referring to Dunlap and Associates (4) Rudy (2) gives UCL R = R ¯¯ + z[R ¯¯ /(DQ)] + 0.829/(DQ) + 1/(2DQ) where z is said to be 2.576 for 1 percent false detection, 1.960 for 5 percent false detection, and so forth. In an appendix attributed to Dietz, it was suggested (3) that the term 0.829/(DQ) be deleted and that UCL R = R ¯¯ + z[R ¯¯ /(DQ)] - DQ/2. These errors should have ended when Baker (5, p. 390) provided the correct expression UCL R = R ¯¯ + z[R ¯¯ /(DQ)] + 1/(2DQ). 54 TRANSPORTATION RESEARCH RECORD 1542 The same can be written more simply as UCL X = R ¯¯ DQ + z(R ¯¯ DQ) + 1 / 2. In this, R ¯¯ DQ is the mean number of accidents for a road section if R ¯¯ was its accident rate and DQ its exposure. If acci- dents are Poisson distributed, (R ¯¯ DQ) is the standard deviation. Thus, UCL X is the sum of what would be expected normally + z standard deviations. The addition of 1 / 2 to this is unimportant. It is important to stress that use of the equation for UCL R is identical to the use of the equation for UCL X . To illustrate, suppose that R ¯¯ = 1.077  10 -6 injury acci- dents/vehicle mile of travel and that a road section recorded 10 injury accidents in 2,932,000 vehicle miles of travel. For this road section the normal number of accidents would be 1.077  10 -6  2,932,000 = 3.16 and thus the standard deviation is 3.16 = 1.77. Clearly, 10  UCL X = 3.17 + 1.77. Plotting exposure as the abscissa and the accident count as the ordinate, this road section is shown in Figure 1 as Point P. Also shown in Figure 1 are 44 addi- tional road sections based on data from Flowers and Griffin (6 ). It is evident which road sections are above the curves for a UCL X . May (9) asks what t, the number of years of accident data to be used, should be. He found that a 13-year average could be ade- quately estimated from 3 years of accident counts. From this May concluded, “There is little to be gained by using a longer study period than three years.” This observation appears to have been entrenched in practice. The motivation must have been to strike a balance and to have sufficient accuracy without using old data that no longer reflect a current situation. However, a sensible choice of t must depend on the magnitude of the average that is being estimated as well as on some knowledge of what makes past accident counts obsolete. This influential guidance is incorrect. Tamburri and Smith (8) introduced the notion of the safety index. This was later incorporated into the practice of black-spot identifi- cation based on the idea that sites with severe accidents deserve prior attention. In principle, each road type was said to have a char- acteristic mix of accident severity. Thus, for example, for a rural two-lane road the mix was 2.9 percent fatal, 43.0 percent injury, and 54.1 percent property damage only (PDO) accidents. They also sug- gested using costs weights by accident severity and road type. If a property damage accident on a rural road was given the weight of 1, fatal and injury accidents on such roads had weights of 95 and 3, respectively. If so, an accident of average severity on a rural two- lane road could be said to be equivalent to 95  0.029 + 3  0.430 + 1  0.541 = 4.6 PDO accidents. Thus, the main idea is to express all accident severities as equivalent PDO (EPDO) accidents. Jorgensen (9) introduced two new ideas. First, E{m} should be estimated by a multivariate model. Second, the ranking should be by the difference between the observed accident frequency of a road section and the expected frequency for such road sections as esti- mated by the multivariate model. The first idea is indeed new inas- much as Norden et al. (1) and successors incorrectly assume that the expected number of accidents for a road section is simply propor- tional to the accident rate. However, the second suggestion is in Identification of Sites with Promise EZRA HAUER Safety Studies Group, Department of Civil Engineering, University of Toronto, Toronto, Ontario M5S 1A4, Canada.