Accident Analysis and Prevention 37 (2005) 870–881

Experimental evaluation of hotspot identification methods

Wen Cheng ∗ , Simon P. Washington 1
Civil Engineering and Engineering Mechanics, University of Arizona, 1209 E. Second St., Tucson, AZ 85721-0072, USA

Received 7 April 2005; received in revised form 10 April 2005; accepted 10 April 2005

Abstract

Identifying crash “hotspots”, “blackspots”, “sites with promise”, or “high risk” locations is standard practice in departments of transportation
throughout the US. The literature is replete with the development and discussion of statistical methods for hotspot identification (HSID).
Theoretical derivations and empirical studies have been used to weigh the benefits of various HSID methods; however, a small number of
studies have used controlled experiments to systematically assess various methods.
Using experimentally derived simulated data—which are argued to be superior to empirical data, three hot spot identification methods
observed in practice are evaluated: simple ranking, confidence interval, and Empirical Bayes. Using simulated data, sites with promise are
known a priori, in contrast to empirical data where high risk sites are not known for certain.
To conduct the evaluation, properties of observed crash data are used to generate simulated crash frequency distributions at hypothetical
sites. A variety of factors is manipulated to simulate a host of ‘real world’ conditions. Various levels of confidence are explored, and false
positives (identifying a safe site as high risk) and false negatives (identifying a high risk site as safe) are compared across methods. Finally,
the effects of crash history duration in the three HSID approaches are assessed.
The results illustrate that the Empirical Bayes technique significantly outperforms ranking and confidence interval techniques (with certain
caveats). As found by others, false positives and negatives are inversely related. Three years of crash history appears, in general, to provide
an appropriate crash history duration.
© 2005 Elsevier Ltd. All rights reserved.

Keywords: Safety; Sites with promise; Empirical Bayesian analysis; Crash history; Hot spot identification

1. Background
The transportation system (consisting of road segments,
intersections, ramps, interchanges, etc.) in general does
not perform homogenously with respect to safety. Not sur-
prisingly, heterogeneity in the driving population, roadside
features, weather, traffic conditions, and design considera-
tions leads to heterogeneity in crash frequencies and rates.
Because of a desire and mandates to provide a safe driving
environment, professionals are charged with identifying and
improving “high risk” locations. A difficulty arises from
the inability to differentiate between sites that are truly high
∗
Corresponding author. Tel.: +1 520 621 4686; fax: +1 520 621 2550.
E-mail addresses: wencheng@u.arizona.edu (W. Cheng),
simonw@engr.arizona.edu (S.P. Washington).
1 Tel.: +1 520 621 4686; fax: +1 520 621 2550.
risk and sites that happen to have experienced random “up”
fluctuation in crashes during a period of observation.
There is a fairly extensive literature focused on methods
for the identification of “blackspots”, “sites with promise”,
“high risk” or “hotspots” (hereafter referred to as HSID for
hot spot identification) covering an exhaustive set of issues.
Some papers address regression to the mean issues, while oth-
ers address crash outcome versus total crash modeling. Some
discuss the application of Bayesian methods, while others try
to make sense of cross-sectional data. There are papers that
discuss the notion of ‘potential accident reduction’ and its
role in hot spot identification, and papers that address issues
surrounding crash severity. Specifically, previous research
(Persaud and Hauer, 1984; Persaud, 1986, 1988; Hauer, 1997)
has shown that methods relying on a simple ranking of crash
counts or crash rates produce large numbers of false posi-
tives – due to the random fluctuation of crashes from year to
year – leading to the attempted remediation of safety prob-

0001-4575/$ – see front matter © 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.aap.2005.04.015
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881
lems at relatively safe locations. In addition, an excessive
number of false negatives allow truly hazardous locations to
escape identification. These errors result in inefficient use
of federal and/or state aid and local government resources
applied for safety improvements. Additionally, the selection
of truly hazardous locations using the method of simple rank-
ing of crash counts is relatively arbitrary, that is, the degree (or
percentage) of difference between “correctable” and “aver-
age” sites is often extremely small and quantitative, not
qualitative.
This observation highlights an important issue, which
warrants further exploration. The separation of “safe” and
“unsafe” sites (intersections, road segments, etc.) is wholly
arbitrary. A binary condition does not describe the differ-
ence between safe and unsafe sites, and in fact the difference
in safety across all sites being compared is continuous in
nature. That is, a safety performance function underlying a
set of sites varies continuously from the lowest crash fre-
quency (or rate) to the highest crash frequency (or rate), the
collection of which results in a probability distribution (or
density) of crashes. In this context it is extremely difficult
to identify clear differences between safe and unsafe sites,
since these two designations will lie alongside one another
in an ordered distribution of all sites, and their difference
in safety (crash counts or rates) may be very small. Some
of the implications of the arbitrary selection of safety and
unsafety of sites are discussed later, but we impart at this
point that identifying high-risk sites is akin to establish-
ing a level of confidence in a statistical hypothesis test. In
practice, the threshold between high risk and non-high-risk
sites is determined by available resources, with the ‘high-
est’ ranked sites being improved until safety resources are
exhausted.
Hauer and Persaud (1984) drew an analogy between the
first stage of identification of black-spots and a sieve, and
discussed how to measure the performances of various meth-
ods of identifying hot-spot sites. On the basis of this study,
Higle and Hecht (1989) conducted a simulation experiment
to evaluate and compare techniques for the identification
of hazardous locations in terms of crash rates. Maher and
Mountain (1988) also use a simulation-based approach to
compare methods, including ranking of sites on the basis of
annual accident totals (AAT) and potential accident reduc-
tion (PAR). Subsequent work by Bauer and Harwood (2000),
Hadayeghi et al. (2003), Miaou and Lord (2003) have shown
that safety performance functions may be curvilinear with
respect to vehicle miles travelled (VMT) and are useful for
assessing the risk of various sites.
The first part of this paper compares alternative HSID
methods and represents a natural continuation of the Higle
and Hecht study, with a number of important differences and
unique contributions. First, crash frequency instead of crash
rate data are used to assess hotspot identification techniques.
Second, the paper compares the performance of simple rank-
ing, classical confidence intervals, and empirical Bayesian
techniques in terms of percent false negatives and positives.
871
Third, several practical empirical crash distributions from the
state of Arizona are selected to represent a realistic range of
‘base’ crash data. Finally, several degrees of crash hetero-
geneity are examined in the experimental evaluation.
In addition to evaluating the performance of various HSID
methods, this paper also evaluates the effect of crash his-
tory duration employed in the three HSID methods. Little
research has been dedicated to exploring the effect of crash
history duration. May (1964) first discussed the issue of how
many years of crash data should be analyzed when deter-
mining the crash-prone locations. He explored the difference
between sorts of average crash counts with t increasing until
13 years. He concluded that the difference diminishes as t
increases while the marginal benefit of increasing t declines.
The “knee” of the curve is said to occur at t = 3 years.
Several researchers (Bauer and Harwood, 2000; Persaud,
1999) also used the same period of crash history records
to address different road safety issues. In this paper, the
effect of crash history duration on HSID results is evaluated
experimentally.
2. Hot spot identification methods
A site (road segments, intersections, interchanges, etc.)
may experience relatively high numbers of crashes due to:
(1) an underlying safety problem, for example, the high level
of traffic exposure or the nature of site; or (2) a random “up”
fluctuation in crash counts during the observation period.
Simply observing unusually high crash counts does not indi-
cate which of the two conditions prevails at the site. Some
researchers have suggested using other measures. For exam-
ple, McGuigan (1981, 1982) proposed the use of potential for
accident reduction, as the difference between the observed
and expected number of crashes at a site given exposure.
Hakkert and Mahalel (1978) proposed that blackspots should
be defined as those sites whose accident frequency is sig-
nificantly higher than expected at some prescribed level of
significance. Mahalel et al. (1982) proposed the road sites
selected for treatment should maximize the expected total
accident reduction by treatment. In light of these various def-
initions, it is necessary to articulate the objective of hot spot
identification as applied in this study.
The objective of hot spot identification (HSID)
is to identify transportation system locations
(road segments, intersections, interchanges,
ramps, etc.) that possess underlying cor-
rectable safety problems, and whose effect will
be revealed through elevated crash frequencies
relative to similar locations.
Two aspects of the previous statement are noteworthy.
First, it is possible to observe an unsafe site that does not
872 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881
reveal elevated crash frequencies – these are termed false neg-
atives. It is also possible to observe elevated crash frequencies
at a relatively safe site – false positives. False positives, if
acted upon, lead to investment of public funds with little to
no safety benefits. False negatives lead to missed opportu-
nities for effective safety investments. As one might expect,
correct determinations include identifying a safe site as “safe”
and an unsafe site as “unsafe”.
For evaluative purposes, a HSID is sought that produces
the smallest proportion of false negatives and false posi-
tives. Hence, percentage of false negatives, false positives,
and overall misidentifications (false positives plus false neg-
atives) are used to compare the performances of three com-
monly implemented techniques: (1) simple ranking of sites,
(2) classically based confidence intervals and (3) empirical
Bayesian methods.
The simple ranking (SR) method is the most straightfor-
ward. Applying this method, a set of locations is ranked in
descending order of crash frequencies (or counts, K), and the
unsafe sites are then identified for engineering evaluation.
This method, for example, is implemented in the current
version of the Arizona Local Government Safety Project
(ALGSP) Model, which is available for local jurisdictions
to conduct HSID (Carey, 2001) in the state of Arizona. Note
that this method does not rely on crash potential (observed
minus expected) but overall crash frequency. It is generally
assumed, however, that similar sites are being compared with
one another.
A second method for HSID is based on classical statisti-
cal confidence intervals (denoted CI) (Laughlin et al., 1975).
Location i is identified as unsafe if the observed crash count
Ki exceeds the observed average of counts of comparison
(similar) locations, µ, with level of confidence equal to δ,
that is, Ki > µ + Kδ S, where S is the standard deviation of the
group of comparison sites. In practice a value of δ is typically
0.90, 0.95, or 0.99, and depends upon the actual situation and
considerations such as the number of sites, amount of safety
investment resources, etc. These values serve as approxima-
tions, since they are borrowed from the normal distribution
function and thus have no special meaning in terms of the
distribution of true crash counts, which typically follows a
Poisson or negative binomial distribution.
The application of empirical Bayesian methods (EB) for
HSID has been developed and applied more recently than
the SR and CI methods. By accounting for both crash his-
tory and expected crashes for similar sites, EB methods have
been shown to offer improved ability to identify “high-risk”
sites. Hauer et al. applied Empirical Bayes’ (EB) methods to
estimate the safety at signalized intersections (Hauer et al.,
1988), Persaud evaluated crash potential of Ontario road sec-
tions (Persaud, 1991) and ranked sites for potential safety
improvements (Persaud, 1999), and Higle and Witkowski
(1988) presented a supplemental EB technique that makes use
of crash rates. Most of these research studies yielded favor-
able results in terms of identifying hotspots, but the range
of conditions were quite small within studies, most relied on
empirical data where actual safety is not known for certain,
and a range of input assumptions was not tested, as is done
in this study.
The empirical Bayesian method rests on the following
logic. Two assumptions about crash occurrence are first
needed, which can be traced back to those of Morin (1967)
and Norden et al. (1956):
Assumption 1. At a given location, crash occurrence obeys
the Poisson probability law. That is, Px|λ denotes the prob-
ability of recording x crashes on a site where their expected
number is λ, where Px|λ = λx e−λ /x!.
Assumption 2. The probability distribution of the λ’s of
the population of sites is gamma distributed, where g(λ) is
denoted as the gamma probability density function, and is a
typically modeled as a function of site covariates.
On the basis of the above assumptions, the probability that
a site selected randomly records x crashes is approximated
by the negative binomial (NB) probability distribution.
In the empirical Bayesian technique, estimation of the
long-term safety of an entity is obtained using two clues, that
is, the historical crash record of the entity and the expected
number of crashes obtained from a safety performance func-
tion for similar sites. If the count of crashes (x) obeys the
Poisson probability law and the distribution of the λ’s in the
reference population is approximated by a Gamma probabil-
ity density function, an estimator of λ, the long term mean
crashes for a site, is given by (Hauer, 1997; Harwood et al.,
2000; Shen and Gan, 2003):
λi = αE{λ} + (1 − α)xi (1)
where α is the weight factor, which can be expressed as fol-
lows:


E{λ}
α= (2)
E{λ} + VAR{λ}
An important issue associated with these two equations is
that the second of the two clues, crash history, significantly
affects the estimate of λ, since longer crash histories tend
to be more stable (in crashes per year) than shorter crash
histories. Thus, different historical crash records yield differ-
ent estimates E{λ} and VAR{λ}, and subsequently different
identification error rates (false positives and false negatives).
Because of its importance in HSID the effect of crash history
duration is examined in this study.
3. Description of simulation experiment
The main objective of this experiment is to quantify and
assess the predictive performance of the SR, CI, and EB HSID
methods and to assess the effect of crash history duration
under a wide range of realistic conditions. Many aspects of
the simulation experiment require careful attention, such as
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881
the reason for simulating data, the determination of sample
sizes, how to characterize crash data, and reliability of tests.
These aspects of the experiment are now discussed in turn.
3.1. Why simulate data when empirical data are
plentiful?
One might ask: why not use empirical data instead of sim-
ulated data to test the performance of HSID methods? The
answer is straightforward. When analyzing real data, i.e.,
crash counts; the analyst never knows a priori which sites
are truly hazardous. This leads to the very difficult situation
of trying to count false positives and negatives without know-
ing which sites are truly safe and ‘unsafe’. In contrast, in a
simulation it is possible to establish a priori sites that are
hazardous and assess whether HSID methods can correctly
identify them. The simulation approach, however, requires
considerable care in ‘constructing’ the crash data so that they
are convincingly similar to empirical crash data.
3.2. Ground rules for simulation experiment
The simulation experiment consists of the following spe-
cific steps, which are described in greater detail following
this summary:
(1) Generate mean crash frequencies from real data. Crash
datasets from Arizona, which represent a range of in
situ crash data, are first obtained. These data are used
to determine various shapes of distributions of crash site
means (λ’s). Gamma distributions are fit to the observed
data to capture the heterogeneity in site crash means.
The gamma-distributed means are denoted TPM’s for
True Poisson Means, and represent the means of crashes
across a collection of sites. To take advantage of ‘large
sample statistical properties’, 1000 TPM’s are generated
for each of the crash distributions analyzed.
(2) From TPM’s generate random Poisson samples. Thirty
independent random numbers are generated from each
TPM. Specifically, for each of the 1000 sites the TPM is
used to randomly generate 30 crash counts that represent
OBSERVED data for 30 different observation periods
(e.g. years).
(3) Evaluate HSID performance. By knowing the true state
of safety for sites (the TPM’s), and having observed data
(the randomly generated Poisson counts) for 30 obser-
vation periods, the performance of HSID methods can
be tested. The following steps are used to conduct the
evaluation:
(a) SR, CI, and EB methods are applied in separate sim-
ulation runs to rank sites for improvement. These
are applied and analyzed column wise (or a single
observation period as shown in Table 1).
(b) For the Bayesian runs, it is assumed that rows (data
across observation periods for the same site) can also
be used to represent the comparison group in order
873
to calculate E(x) and VAR(x). This implies that the
analyst has accommodated for covariates and is able
to estimate an expected value for a site that accounts
for things such as exposure, geometrics, etc.
(c) For the various hot spot thresholds, false positives,
false negatives, and total misidentifications in per-
cent are computed.
(4) Evaluate effect of crash history duration. In the CI, SR,
and EB methods the analyst must decide how long a
history to use for calculations. In this experiment the
effects of various crash histories (1 year through 10 years
of data) on performance are evaluated.
3.3. Generating mean crash frequencies from empirical
data
As stated previously, observed crash data are used to
inform the simulation of the TPM’s. Herein, the 6-years (Jan-
uary 1995 through December 2000) of crash counts from
intersections in Apache, Gila, Graham, Lapaz, Pima, and
Santacruz counties in the state of Arizona are used. The cumu-
lative distributions (i.e., cumulative percent) of crash counts
for the different sites are shown in Fig. 1 (1 plot for each of
the six counties). Three types of characteristically different
underlying cumulative distributions of TPM’s were observed:
an approximately exponential shape (denoted E), an approx-
imately linear shape (denoted L), and an approximately sig-
moidal shape (denoted S). These three distributional shapes
are meant to reflect different types of facilities that have
varying safety performance functions, where the safety per-
formance function characterizes the collection of sites.
In addition, two levels of heterogeneity in crash counts
were observed: low heterogeneity (denoted 1) where the
range in observed crash counts is less than 20 crashes (differ-
ence between highest and lowest crash count in a single obser-
vation period), and high heterogeneity (denoted 2) where the
range is in excess of 50 crashes. These two groups are meant
to represent low exposure and high exposure sites.
The cumulative distributions used to generate the TPM’s
are labeled as E1, E2, L1, L2, S1, and S2, and are shown
in Fig. 1. For example, E2 represents a collection of sites
with an exponentially shaped crash distribution with high
heterogeneity in TPM’s. These six data sets, which by good
luck provided these distinctive distributions, were selected
from various jurisdictions within Arizona to try to represent
the range of underlying characteristics related to true acci-
dent count distributions, with the intent of making the results
gained from this experiment applicable across a verity of typ-
ical situations.
3.4. Generation of random Poisson samples from TPM’s
The empirical cumulative TPM’s shown in Fig. 1 represent
the data required to support Assumption 1. Using these data,
observed crash counts were randomly generated to repre-
sent observed data for a given observation period. To provide
874 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881

Table 1
Simulated crash counts for 30 sites and 16 observation periods
Site TPM Observation period
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 4 5 1 4 1 2 7 4 3 4 4 2 1 1 5 5 6
2 8 5 9 8 6 8 4 9 9 5 4 8 8 9 9 13 8
3 8 12 7 10 5 5 7 11 8 8 8 11 6 6 7 8 7
4 9 12 9 10 16 8 12 7 9 11 8 10 8 16 11 6 8
5 9 10 13 12 8 9 6 12 10 9 9 4 5 12 11 11 4
6 10 15 4 6 10 4 17 6 11 12 7 10 10 15 6 17 10
7 10 8 5 10 8 13 10 11 7 12 10 8 9 9 6 9 10
8 10 7 8 11 14 10 12 7 11 12 11 12 13 7 7 7 11
9 12 13 17 8 14 12 10 16 10 7 15 17 9 11 15 14 15
10 12 10 9 13 13 6 12 18 11 15 12 12 12 13 12 13 9
11 12 9 10 10 14 15 12 7 14 6 12 11 19 9 17 10 18
12 12 11 14 14 9 16 7 15 3 10 13 9 11 7 2 12 14
13 12 15 15 16 13 8 12 13 16 16 12 15 11 15 12 14 9
14 12 14 10 10 11 15 15 12 13 14 15 13 14 11 13 17 19
15 12 11 12 12 8 12 13 12 7 9 11 9 9 9 12 4 9
16 13 8 17 13 8 12 11 17 15 16 13 12 15 16 12 14 19
17 13 9 13 16 16 11 8 6 18 12 8 7 11 12 12 17 15
18 13 10 18 15 16 10 15 10 16 17 10 6 8 8 10 13 6
19 13 14 13 17 11 6 11 18 15 11 17 16 19 13 11 15 14
20 13 7 4 13 11 12 10 17 19 6 7 12 15 7 15 14 12
21 14 16 17 12 18 13 17 12 11 7 13 15 10 18 14 17 19
22 15 15 18 21 15 15 14 13 21 14 13 20 13 12 19 16 16
23 15 11 13 16 12 12 16 10 16 19 20 21 16 13 19 11 16
24 15 9 16 16 11 14 12 15 18 11 16 14 29 11 12 19 14
25 16 18 12 15 9 19 18 14 11 19 15 18 14 18 18 14 20
26 17 22 10 19 12 15 19 18 10 11 17 20 16 15 11 10 15
27 18 14 21 9 19 16 17 19 18 18 14 16 28 19 18 19 10
28 18 8 20 19 5 16 18 20 28 16 17 19 14 15 14 18 15
29 19 26 19 18 21 17 29 12 22 25 15 23 11 19 20 15 24
30 20 22 18 23 21 23 19 26 22 16 20 19 15 14 19 13 15
Site = number of site, e.g. intersection, road segment, etc.; TPM = true underlying safety of site or Poisson mean; simulated data = observed crash count in
observation period; shaded cells represent ‘truly hazardous’ locations (sites 19 and 20).

sufficient sample sizes for statistical comparisons, gamma After TPM’s are simulated (the crash means across sites
distributions of TPM’s are fitted to the six datasets, then which reflects the true and unknown safety performance func-
1000 TPM’s are simulated. Fitting gamma distributions to tion) the next step is to generate observed crash counts for
a given sequence of data was implemented through the soft- the sites. These counts represent the observed crash counts
ware package Arena 7.0 (Kelton et al., 2003). The summary across observation periods for a particular site (where its’ true
of the fittings is shown in Table 2. Differences in theoret- safety, the TPM, is known). It is well established that crash
ical curves and empirical data are expected, since the data counts that fluctuate across observation periods is a result of
represent the sum of an underlying (and unknown) safety the randomness inherent in the underlying crash process and
performance function and random crash counts. is well approximated by a Poisson process (see for example
Lord et al., 2004). To represent this natural fluctuation, a ran-
dom sample of 30 observation periods (e.g. months, years,
Table 2 etc.) associated with each location is simulated using a ran-
Summary of Gamma fittings of six datasets dom number generator using the TPM’s defined by the fitted
Data Fitting expression Square error Test p-value distributions in Fig. 1. A small snapshot of the data generated
set statistic by this simulation is shown in Table 1.
E1 0.5 + Gamm(3.79,1.75) 0.022344 26 <0.005 Table 1 shows 16 simulated observations periods for 30
E2 1.5 + Gamm(15.9,1.7) 0.011836 13.4 0.0385 sites, and is perhaps the best way to illustrate how the sim-
L1 0.5 + Gamm(4.31,1.71) 0.038173 11.1 0.0119
ulation was conducted. The far left column shows 30 sites
L2 3.5 + Gamm(13.4,2.27) 0.020052 8.2 0.16
S1 0.5 + Gamm(2,4.3) 0.014903 33.5 <0.005 by number. The next column, labeled TPM, is the true state
S2 0.5 + Gamm(9.06,2.57) 0.013211 23 <0.005 of nature for a particular site, and represents the usually
E—exponential shape; L—linear shape; S—sigmoidal shape; 1—low het- unknown underlying safety. The remaining 16 columns rep-
erogeneity of crash counts; 2—high heterogeneity of crash counts. resent 16 different observation periods – years of crash data
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881
Fig. 1. Cumulative distributions of TPM’s underlying simulated crash data.
– that could possibly arise from the TPM’s in column 2.
The data, in aggregate (across sites and observation periods),
are negative binomial distributed, but are Poisson distributed
within site.
To illustrate how high risk sites might be identified, the two
sites with 19 or more crashes per observation period may be
identified a priori as hazardous, since the TPM’s reflect the
true underlying state of nature. The two sites in the shaded
cells are truly high-risk sites whereas the 28 sites above the
shaded area are ‘safe’. Using a simple ranking scheme, in
any given observation period, say observation period 5, the
observed number of truly hazardous sites that recorded 19 or
more crashes was 2 sites out of 30, where one was a truly
hazardous site (site 30) and one was not (site 25, a false pos-
itive). In observation period 5 there was also a false negative,
since truly hazardous site 29 revealed only 17 crashes. Using
a large number of observation periods in this fashion will
yield statistical estimates of false positives and negatives.
4. Performance evaluation results for HSID methods
Establishing fair and realistic comparisons among the dif-
ferent HSID methods is paramount. One consideration in
875
this regard is the use of δ—the cutoff level used to establish
hazardous locations. Three values of δ are employed in the
evaluations, 0.90, 0.95, and 0.99 corresponding to the top 10,
5, and 1% of all sites identified as ‘hazardous’ respectively.
In practice δ corresponds with the availability of resources
for remediation.
The parameters of the simulation experiment in their
entirety now include: shapes of the TPM’s (E, S, and L),
levels of heterogeneity in the TPM’s (1 and 2), and lev-
els of δ (0.90, 0.95, and 0.99). Three HSID methods are
assessed, simple ranking (SR), classical confidence intervals
(CI), and empirical Bayesian methods (EB). The evaluation
criteria include percent of false positives (FP), percent of false
negatives (FN), and sum total percent of false positives and
false negatives, called false identifications (FI). The true state
of nature (TPM) is simulated for 1000 sites for each of the
six crash distributions and each site is “observed” during 30
observation periods.
To conduct this simulation experiment the following steps
were taken:
1. All the TPM cumulative distributions are divided into truly
hazardous locations and non-hazardous locations, using
thresholds of 0.90, 0.95, and 0.99 to represent different
876 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881

data separation thresholds. This step results in three “crit-
ical” crash count threshold values, CC0.90 , CC0.95 , and
CC0.99 for each combination of cumulative TPM shape
and heterogeneity level.
2. The three different HSID methods are used to identify hot
spots using the simulated data. Specifically, the SR method
simply ranks observed frequencies, the CI method uses
the entire sample mean and standard deviation to deter-
mine confidence intervals for ranking, and the EB method
uses a weighted average of crash history and observed
frequency using gamma distribution parameters to rank
sites.
3. Simulated crash data are then compared to the values of
CC0.90 , CC0.95 , and CC0.99 . For truly hazardous sites; if
the randomly generated crash counts are lower than the
values CC0.90 , CC0.95 , and CC0.99 , then false negatives
are produced. Similarly, for non-hazardous sites, when the
simulated crash counts are larger than the values CC0.90 ,
CC0.95 , and CC0.99 , false positives are generated. The
number of false identifications is the sum total of the num-
ber of false negatives and positives.
4. To make the three performance metrics comparable across
simulations, the percentage of false negatives, false pos-
itives, and false identifications are calculated. The per-
centage of false negatives is the number of simulated
false negatives divided by the simulated truly safe sites,
the percentage of the false positives is the number of
false positives divided by the truly hazardous locations,
and the percentage of false identifications is obtained by
dividing the sum of false negatives and false positives by
the total number of randomly generated data locations
(1000).
5. Finally, the percentage of false positives, false negatives,
and false identifications across simulation conditions are
tallied and reported.
Tables 3 and 4 summarize the results of the errors (FNs,
FPs, and FIs) produced under the simulated conditions.
Table 3 presents the results when heterogeneity of crash
counts is low (1), while Table 4 presents the results when
heterogeneity is high (2). Critical crash count threshold val-
ues increase from left to right in both of the tables.
For low heterogeneity and high heterogeneity simulations,
the trends of percent errors with increasing δ are consistent,
however, the percentage of errors for low heterogeneity are
much higher than those for high heterogeneity. The major
reason is that low heterogeneity in crash counts results in rel-
atively small standard deviations of crashes when compared
with the other datasets, which in turn makes it more difficult
to identify hazardous locations. On the contrary, it is easy to
identify hot spots when the corresponding crash counts are
greatly dispersed, particularly when dispersion is large in the
upper most crash count deciles.
Another prominent characteristic illustrated in the tables is
that the percentage of false negatives decreases with increas-
ing δ for the three HSID methods. In most cases the per-
centage of false negatives is substantially smaller for the EB
method. The fairly complicated explanation for this is as fol-
lows. The threshold value divides the top ‘outlying’ crash
counts from the remainder of the data, either the top 10, 5, or
1% of observed counts. By definition these counts are more
likely to suffer from regression to the mean in a subsequent
observation period than counts around the TPM. Thus the
crash history of the top x% of crash counts act to reduce the
effect of the current crash count x when ranking these sites.
As a result sites that suffer less from regression to the mean
get ranked higher in the list—sites that ordinarily would have
been ranked as false negatives.
Conversely, the percentage of false positives increases
with increasing δ for the three HSID methods (except for
the δ of 0.95 for L1 and L2). This result suggests that stricter

Table 3
Percent errors for low heterogeneity in crash counts
Percent errors: low heterogeneity

δ 0.9 0.95 0.99

Method CI SR EB CI SR EB CI SR EB

E FN 2.49 3.55 2.40 1.54 2.09 1.41 0.63 0.55 0.38

FP 62.76 31.97 21.63 82.47 39.73 26.87 114.32 54.00 37.67
FI 7.17 6.39 4.33 5.31 3.97 2.69 2.46 1.08 0.75

L FN 2.21 4.44 2.91 1.39 2.40 1.73 0.15 0.62 0.45

FP 106.14 39.97 26.20 65.24 45.67 32.80 431.62 61.00 45.00
FI 8.75 7.99 5.24 3.62 4.57 3.28 2.10 1.22 0.90

S FN 0.54 6.53 5.28 0.21 3.48 2.90 0.00 0.81 0.73

FP 753.44 58.73 47.50 1251.33 66.20 55.13 NA 80.33 72.33
FI 10.03 11.75 9.50 6.46 6.62 5.51 1.91 1.61 1.45
FN—false negatives; FP—false positives; FI—false identifications; CI—confidence interval; SR —simple ranking; EB—empirical Bayesian; E—exponential
shape; L—linear shape; S—sigmoidal shape. In the table, the reason that some FPs can exceed 100% is due to non-normality of the distribution and setting of
threshold, and in these cases, the CI method identifies more hazardous locations than truly exist. For the same reason, the existing of “NA” in the table is due
to zero truly hazardous locations identified by confidence analysis. The shaded cells show the lowest identification error rate.
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881 877

Table 4
Percent errors for high heterogeneity in crash counts
Percent errors: high heterogeneity
δ 0.9 0.95 0.99

Method CI SR EB CI SR EB CI SR EB

E FN 1.78 2.09 1.13 1.33 1.33 0.86 0.39 0.26 0.17

FP 24.37 18.77 10.13 32.56 25.33 16.40 57.07 26.00 16.67
FI 4.13 3.75 2.03 3.34 2.53 1.64 1.54 0.52 0.33
L FN 1.89 2.55 1.57 1.50 1.43 0.91 0.44 0.37 0.23
FP 36.33 22.93 14.13 32.20 27.20 17.33 45.22 36.67 22.67
FI 5.14 4.59 2.83 3.40 2.72 1.73 1.29 0.73 0.45
S FN 2.16 2.73 1.74 1.17 1.31 0.71 0.47 0.26 0.12
FP 34.80 24.53 15.67 41.08 24.87 13.47 38.37 25.33 12.33
FI 5.16 4.91 3.13 3.31 2.49 1.35 1.32 0.51 0.25

FN—false negatives; FP—false positives; FI—false identifications; CI—confidence interval; SR—simple ranking; EB—empirical Bayesian; E—exponential
shape; L—linear shape; S—sigmoidal shape. The shaded cells show the lowest identification error rate.

identification criteria (higher δ) will result in fewer failures to
select truly hazardous sites for remedy, but will lead to a larger
number of non-hazardous locations identified as hazardous.
In summary, the percent of false positives increases with the
rising of thresholds, whereas the percent false negatives and
false identifications decrease with the rising of thresholds.
The results in almost all simulation scenarios reveal similar
trends.
The three identification methods also perform differently.
Compared to the other two more traditional methods, the EB
method yields fewer false negatives and false positives in
most simulation scenarios (see shaded ‘best’ results for each
of the simulations in the table). In many cases the EB method
reduced false negatives by about 50% compared to CI and SR
methods, while the reduction in false positives ranges from 30
to 50%. That is, the EB technique is more efficient in iden-
tifying sites that require further analysis and/or inspection.
As for the CI and SR methods, there is not a significant per-
formance difference between them; however, the CI method
slightly outperforms the SR method for low-heterogeneity in
crash count situations.
5. Results for assessment of crash history duration
To assess crash history duration effects, error rates asso-
ciated with various “t” years of crash history are compared.
All simulation conditions are repeated for the crash history
analysis. An experimental enhancement lies in how to use
the different periods of data. A snapshot of data in Table 1
are used again to demonstrate the analysis procedure used
here. The ith column of data is assumed to represent the ith
current year of crash data for all sites. Consider conducting
an EB analysis: it is known that for a given t-year period,
Eqs. (1) and (2) are used for each site to compute the corre-
sponding expected crash counts. For a t-year crash history,
average crash counts per year should be used in these equa-
tions. For example, in the 4th year or observation period, the
crash count for site 28 is 13 crashes per year (average of the
first four observation periods), while E{λ} = 15.75 crashes
(row or comparison site average), VAR{λ} = 4.14 crashes2
(row or comparison site variance), and the formula weight
α = 0.792. The resulting expected crash count associated with
site 28 when using the first 4-year data is 15.2 crashes. For
the 16 different observation periods (shown in the table) we
can generate 13 expected crash counts associated with site 28
using the 4-year crash history record. This same procedure
is applied for varying crash histories. Furthermore, the same
analysis approach is applied to assess the effect of crash his-
tory using the SR and CI methods. Due to a large amount
of iterative computations in this experiment, computer code
was written to calculate the various identification error rates
associated with different periods of crash data.
In theory, as t increases, the average expected crash counts
will converge to the site’s TPM (recall that the simulated
data follow the Poisson distribution) and the correspond-
ing identification error rate converges to zero. However, in
practice, as t increases each site will suffer from changes
by influential factors such as traffic volumes, driver popu-
lation, maintenance activities, surrounding changes to land
use, weather fluctuations, etc., and thus longer crash history
periods generally are associated with increasingly less stable
safety performance functions with time. In contrast, if a short
period of data is used, the count suffers from random fluc-
tuations and/or regression to the mean bias. Consequently, a
trade-off is sought between a study period that is short enough
to represent the current conditions and long enough to reflect
the true expected crash count of a site. In this experiment,
various identification rates are plotted versus t years of crash
history duration, and a “knee” in this curve is expected to
serve as guidance for period selection.
To consider up to 10 years of crash history, the 30 simu-
lated observation periods are divided into three groups: the
first 10 columns of data belonging to group 1, the 11th to
20th column of data belonging to group 2, and the last 10
columns of data belonging to group 3. This grouping effec-
878 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881

tively increases the size of the sample used for the analysis
and eliminates consideration of crash histories longer than
10 years, which practically represents an upper limit.
Generally, a plot of the identification error rate and t plot-
ted against each other will reveal fluctuations along the curve
due to random fluctuations characteristic of stochastic data.
To quickly identify the initial “warm-up” period – the period
prior to the knee of the curve – the moving average method
is utilized.
The moving average Ȳi (w) (where w, the window size) of
random observations is defined as follows:
y
 i−w + · · · + yi + · · · + yi+w

 , Fig. 2. Moving averages vs. original statistic.

 2w + 1

 i = w + 1, . . . , m − w
Ȳi (w) = y + · · · + y + · · · + y (3) tional crash history data yields diminishing improvement in

 1 i 2i−1

 , the percentage of false negatives.

 2i − 1
 To assess the effect of crash duration history, 486 simula-
i = 1, . . . , w
tion scenarios (three identification methods, three distribution
In this experiment, a window size of 1 is selected. Details of shapes, low and high heterogeneity for crash counts, three
the moving average method can be found in Simulation with threshold values for truly hazardous locations, three kinds
ARENA (Kelton et al., 2003). of false identifications, or, FN, FP, FI and three groups) were
The moving average effectively smooths the statistical conducted and analyzed to yield the corresponding best study
fluctuations in observations (yi ) and illustrates more clearly duration of crash history—where “best” refers to the study
the “warm-up” period. As illustrated in Fig. 2, it is difficult period associated with the knee of the curve. Table 5 shows
to detect the “knee” of the curve due for the unsmoothed data the identification error rates for the EB method for group 1
(original plot), due primarily to the existence of two outliers data.
(the point when t = 4 and t = 6). However, it is relatively easy The frequencies of best study periods for different confi-
to detect the knee of the curve in the plot of moving aver- dence levels and the total number of sites is shown in Fig. 3,
ages; it appears that a t of 5 years is close to the ‘knee’ of where the results from three groups of data mentioned previ-
the curve. The knee of the curve is the point at which addi- ously are aggregated. Moreover, the cumulative distribution

Table 5
The identification error rates of different crash histories for group 1
1 2 3 4 5 6 7 8 9 10

E1 FN 2.47 1.73 1.53 1.51 1.44 1.31 1.31 1.19 1.22 1.22
FP 22.20 15.56 13.75 13.57 13.00 11.80 11.75 10.67 11.00 11.00
FI 4.44 3.11 2.75 2.71 2.60 2.36 2.35 2.13 2.20 2.20

E2 FN 1.19 0.83 0.67 0.59 0.54 0.60 0.47 0.48 0.50 0.56
FP 10.70 7.44 6.00 5.29 4.83 5.40 4.25 4.33 4.50 5.00
FI 2.14 1.49 1.20 1.06 0.97 1.08 0.85 0.87 0.90 1.00

L1 FN 3.03 2.40 2.13 2.00 1.85 1.87 1.78 1.78 1.72 1.67
FP 27.30 21.56 19.13 18.00 16.67 16.80 16.00 16.00 15.50 15.00
FI 5.46 4.31 3.83 3.60 3.33 3.36 3.20 3.20 3.10 3.00

L2 FN 1.87 1.41 1.36 1.29 1.28 1.20 1.08 1.04 1.06 1.11
FP 16.80 12.67 12.25 11.57 11.50 10.80 9.75 9.33 9.50 10.00
FI 3.36 2.53 2.45 2.31 2.30 2.16 1.95 1.87 1.90 2.00

S1 FN 5.39 4.75 4.33 4.11 4.02 3.91 3.92 3.81 3.61 3.33
FP 48.50 42.78 39.00 37.00 36.17 35.20 35.25 34.33 32.50 30.00
FI 9.70 8.56 7.80 7.40 7.23 7.04 7.05 6.87 6.50 6.00

S2 FN 2.08 1.81 1.65 1.59 1.56 1.51 1.08 1.41 1.33 1.44
FP 18.70 16.33 14.88 14.29 14.00 13.60 9.75 12.67 12.00 13.00
FI 3.74 3.27 2.98 2.86 2.80 2.72 1.95 2.53 2.40 2.60

EB method and δ = 90%. FN—false negatives; FP—false positives; FI—false identifications; CI—confidence interval; SR—simple ranking; EB—empirical
Bayesian; E—exponential shape; L—linear shape; S—sigmoidal shape. The shaded cells shows that where the “knees” are found.
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881 879

Fig. 3. The frequencies of t-year which is the “knee” of the curve for three confidence levels and the total.

of the sum of the three confidence levels illustrated in Fig. 3
is shown in Fig. 4.
Figs. 3 and 4 show that among 486 simulation scenarios, a
3-year crash history represented the largest portion of “best”
study periods of crash history, and 3 through 6 years make
up almost 90% of all the optimum t-years. Hence, consider-
ing the trade-off between the long and short history record,
if there is no significant physical change in the sites being
examined and a longer history record can be obtained, it is
suggested that 3 or more years be used (up to 6 years). In
contrast, 3-years of crash history data represents the ‘short-
est’ period of time that should be used and which achieves a

Table 6
Percent errors for low heterogeneity in crash counts (3 years data)
Percent errors: low heterogeneity

δ 0.9 0.95 0.99

Method CI SR EB CI SR EB CI SR EB

E FN 2.02 2.32 1.53 1.36 1.34 0.82 0.89 0.40 0.25

FP 28.06 20.88 13.75 38.60 25.50 15.50 48.56 40.00 25.00
FI 4.68 4.18 2.75 3.69 2.55 1.55 2.13 0.80 0.50
L FN 2.56 2.75 2.13 1.69 1.72 1.25 0.47 0.51 0.40
FP 33.16 24.75 19.13 50.00 32.75 23.75 91.07 50.00 40.00
FI 5.56 4.95 3.83 4.33 3.28 2.54 0.14 0.67 0.53

S FN 1.10 4.88 4.33 0.68 2.88 2.54 0.14 0.67 0.53

FP 228.21 43.88 39.00 239.38 54.75 48.25 362.16 66.25 52.50
FI 9.05 8.78 7.80 5.45 5.48 4.83 1.81 1.33 1.05
FN—false negatives; FP—false positives; FI—false identifications; CI—confidence interval; SR —simple ranking; EB—empirical Bayesian; E—exponential
shape; L—linear shape; S—sigmoidal shape. In the table, the reason that some FPs can exceed 100% is due to non-normality of the distribution and setting of
threshold, and in these cases, the CI method identifies more hazardous locations than truly exist. For the same reason, the existing of “NA” in the table is due
to zero truly hazardous locations identified by confidence analysis. The shaded cells show the lowest identification error rate.
880 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881

Table 7
Percent errors for low heterogeneity in crash counts (3 years data)
Percent errors: high heterogeneity
␦ 0.9 0.95 0.99

Method CI SR EB CI SR EB CI SR EB

E FN 1.08 1.28 0.67 0.96 0.95 0.71 0.24 0.14 0.10

FP 13.96 11.50 6.00 15.32 18.00 13.50 34.66 13.75 10.00
FI 2.51 2.30 1.20 1.98 1.80 1.35 1.00 0.28 0.20
L FN 1.72 1.63 1.36 1.19 0.96 0.87 0.41 0.21 0.20
FP 14.37 14.63 12.25 15.07 18.25 16.50 20.11 21.25 18.25
FI 3.08 2.93 2.45 2.14 1.83 1.65 0.86 0.43 0.38
S FN 2.10 2.04 1.65 0.70 0.66 0.55 0.40 0.15 0.10
FP 18.01 18.38 14.88 20.83 12.50 10.50 21.03 15.00 10.00
FI 3.73 3.68 2.98 1.85 1.25 1.05 0.90 0.30 0.20
FN—false negatives; FP—false positives; FI—false identifications; CI—confidence interval; SR —simple ranking; EB—empirical Bayesian; E—exponential
shape; L—linear shape; S—sigmoidal shape. The shaded cells show the lowest identification error rate.

uate the performance of simple ranking, confidence interval,

Fig. 4. The cumulative distribution of various optimal study periods.
significant benefit of crash history (under most general con-
ditions). Crash histories of 1 and 2 years provided relatively
little benefit under the range of conditions tested.
Tables 6 and 7 illustrate the improvement in identifica-
tion performance results after using 3-year crash history data
(in contrast to Tables 3 and 4 which use 1 year of crash
data). Comparison of these tables reveals that using 3 years
of crash history data results in significant improvements in
error rates for all three methods, CI, SR, and EB. Moreover,
improvements are seen across most scenarios (except for two
false negatives for method CI)—with the EB method show-
ing 10–20% reductions in errors on average. While the EB
method still reveals itself as the superior method, the SR
and CI methods benefit disproportionately, more, on aver-
age, from using longer crash histories—with reductions in
errors ranging from 15 to 50%.
6. Conclusions and recommendations
Real crash data from various intersections from six coun-
ties in the State of Arizona were used to simulate crash data
with known parameters. The simulated data enable a priori
knowledge regarding the true (and usually unknown) safety
of simulated sites. The simulated data are then used to eval-
and empirical Bayesian high risk site identification methods
under a variety of realistic conditions. The evaluation criteria
are false positives, false negatives, false identifications, and
diminishing returns of crash history duration, t. The study dif-
fers from the vast majority of past research which has relied
upon empirical data (where underlying safety of sites is not
known a priori), where crash rates were used, or where the
effect of crash history has received little attention.
Upon examination of the simulation results described in
detail in this paper, the following conclusions are and recom-
mendations are made:
(1) Empirical Bayesian methods in general outperform the
other two relatively conventional HSID methods. Under
a range of practical conditions, the EB method offers
50% reductions in the percentages of false positives
and false negatives compared to CI and SR methods.
The EB analysis benefits, however, are contingent upon
reliable and accurate safety performance functions for
predicting ‘expected’ safety of comparison sites, which
requires good geometric, traffic, and crash data, and is a
caveat for the results observed in this investigation. It is
strongly recommended that EB methods be incorporated
into mainstream practice by managers of road safety who
currently may use CI or SR methods.
(2) In low crash count heterogeneity situations the benefits
of the EB methods are relatively less pronounced (com-
pared to high crash count variability). That is, when the
observed differences in crashes between ‘high-risk’ and
‘safe’ sites is relatively small, the EB method offers only
minor improvement compared to the SR and CI meth-
ods. This might suggest that municipalities that manage
safety on systems with relatively few crashes and low
exposure may not experience significant improvements
in performance by changing analysis platforms from SR
and CI methods to the EB method.
(3) The analysis of crash history suggests that a 3-year crash
history is the ‘optimum’ crash history, and up to 6 years of
 W. Cheng, S.P. Washington / Accident Analysis and Prevention 37 (2005) 870–881
crash history is generally better than shorter histories. It
is recommended that not less than 3 years of crash history
duration be used in a HSID analysis, and the most recent
6-year crash history record may be used if few substantive
changes at the site during this period occurred.
(4) Finally, a drastic improvement of truly identifying
hotspots between the SR method (which is still used in
practice) with 1 year of crash data and the EB method
with 3 years of crash data is possible. For example, the
percent of false negatives and false positives associated
with the latter ranges between 25 and 50% less than those
associated with the former. A significant improvement of
decreasing false identification rates in the results from SR
and CI methods is also possible by including 3–6 years
of crash data, but these methods are still outperformed
by the EB method.
Although the research here reflects an improved under-
standing of how various HSID perform, further work is still
needed. The results here depend on reliable and accurate
safety performance functions which are not always available.
Thus, it is not known how beneficial EB methods perform rel-
ative to other methods when safety performance functions are
inaccurate. Simulated data also suffer from lack of realism
encountered in ‘uncontrolled’ observational settings. Factors
such as changes in weather, road users, enforcement, and
special events (among others) affect crash counts and in par-
ticular the crash history analysis to favor shorter periods.
These factors also may lead to long-term trends in crashes,
which have not been considered in this analysis. In addition,
crash data that do not follow a Poisson–Gamma distribu-
tion (negative binomial) may perform differently. Finally, it
may be possible to improve upon the EB method and further
reduce false positives and negatives.
Acknowledgements
This paper is based on the research performed through
contact with the Arizona Department of Transportation,
whose supply of the Arizona LGSP Model and crash data is
gratefully acknowledged, along with support of the research.
The contents are the sole responsibility of the authors.
References
Bauer, K.M., Harwood, D., 2000. Statistical models of at-grade intersec-
tion accidents—addendum, Publication FHWA-RD-99-094.
Carey, J., 2001. Arizona local government safety project (LGSP) analysis
model (Final Report 504). Phoenix, AZ.
Hadayeghi, A., Shalaby, A.S., Persaud, B.N., 2003. Macrolevel accident
prediction models for evaluating safety of urban transportation sys-
tems. Transport. Res. Rec. 1840, 87–95.
881
Hakkert, A.S., Mahalel, D., 1978. Estimating the number of accidents at
intersections from a knowledge of the traffic flow on the approaches.
Accident Anal. Prevent. 10, 69–79.
Harwood, D.W., Council, F.M., Hauer, E., Hughes, W.E., Vogt, A.,
2000. Prediction of the expected safety performance of rural two-lane
highways, Publication FHWA-RD-99-207. FHWA, US Department of
Transportation.
Hauer, E., Persaud, B.N., 1984. Problem of identifying hazardous loca-
tions using accident data. Transport. Res. Rec. 975, 36–43.
Hauer, E., NG, J.C.N., Lovell, J., 1988. Estimation of safety at signalized
intersections. Transport. Res. Rec. 1185, 48–61.
Hauer, E., 1997. Observational Before-After Studies in Road Safety. Perg-
amon, Tarrytown, NY.
Higle, J.L., Witkowski, J.M., 1988. Bayesian identification of hazardous
locations. Transport. Res. Rec. 1185, 24–36.
Higle, J.L., Hecht, M.B., 1989. A comparison of techniques for the iden-
tification of hazardous locations. Transport. Res. Rec. 1238, 10–19.
Kelton, W.D., Sadowski, R.P., Sturrock, D.T., 2003. Simulation with
Arena. The McGraw-Hill Companies, Inc., NY.
Laughlin, J.C., Hauer, L.E., Hall, J.W., Clough, D.R., 1975. NCHRP
Report 162: methods for evaluating highway safety improvements.
National Research Council, Washington, D.C.
Lord, Dominique, S. Washington, and J. Ivan., 2004. Poisson,
Poisson–Gamma, and Zero-Inflated Regression Models of Motor Vehi-
cle Crashes: Balancing Statistical Fit and Theory. Accident Analysis
and Prevention, Pergamon Press/Elsevier Science.
Mahalel, D., Hakkert, A.S., Prashker, J.W., 1982. A system for the allo-
cation of safety resources on a road network. Accident Anal. Prevent.
14 (1), 45–56.
Maher, M.J., Mountain, L.J., 1988. The identification of accident
blackspots: a comparison of current methods. Accident Anal. Prevent.
20 (2), 143–151.
May, J.F., 1964. A determination of accident prone location. Traffic Eng.
34 (5), 21–27.
McGuigan, D.R.D., 1981. The use of relationships between road accidents
and traffic flow in “blackspot” identification. Traffic Eng. Contr. 22
(8/9), 448–451, 453.
McGuigan, D.R.D., 1982. Non-junction accident rates and their use in
“black-spot” identification. Traffic Eng. Contr. 23 (2), 60–65.
Miaou, S., Lord, D., 2003. Modeling traffic crash-flow relationships for
intersections: dispersion parameter, function form, and Bayes versus
empirical Bayes methods. Transport. Res. Rec. 1840, 31–40.
Morin, A., 1967. Application of statistical concepts to accident data. High-
way Res. Rec. 187, 72–79.
Norden, M., Orlansky, J., Jacobs, H., 1956. Application of statistical
quality-control techniques to analysis of highway-accident data. High-
way Res. Rec. 117, 17–31.
Persaud, B.N., Hauer, E., 1984. Comparison of two methods for de-
biasing before-and after accident studies. Transport. Res. Rec. 975,
43–49.
Persaud, B.N., 1986. Safety migration, the influence of traffic volumes,
and other issues in evaluating safety effectiveness. Transport. Res.
Rec. 1086, 33–41.
Persaud, B.N., 1988. Do traffic signals affect safety? Some methodolog-
ical issues. Transport. Res. Rec. 1185, 37–46.
Persaud, B.N., 1991. Estimating accident potential of Ontario road sec-
tions. Transport. Res. Rec. 1327, 47–53.
Persaud, B.N., 1999. Empirical Bayes procedure for ranking sites for
safety investigation by potential for safety improvement. Transport.
Res. Rec. 1665, 7–12.
Shen, J., Gan, A., 2003. Development of crash reduction factors: methods,
problems, and research needs. Transport. Res. Rec. 1840, 50–56.