Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Safety Impacts of Differential Speed Limits for Trucks

and Passenger Cars on Rural Interstate Highways:


A Modified Empirical Bayes Approach
Nicholas J. Garber1; John S. Miller2; Xin Sun3; and Bo Yuan4

Abstract: To compare the safety effects of a uniform speed limit 共USL兲 for all vehicles as opposed to a differential speed limit 共DSL兲
for cars and heavy trucks, detailed crash data for six states were obtained for rural interstate highways for the period 1991–2000. The
states were categorized into four policy groups based on the speed limit type employed during that decade: maintenance of a uniform
limit, maintenance of a differential limit, a change from a uniform to a differential limit, and a change from a differential to a uniform
limit. A modified empirical Bayes framework was used to evaluate crash frequency changes without presuming a constant relationship
between crashes and traffic volume. Aggregate results showed no consistent safety effects of DSL as opposed to USL. The reason for this
finding is that within each state, the modified empirical Bayes methodology suggested that crash risk increased regardless of speed limit
policy. The crash estimation models 共CEMs兲 that were developed for each state varied substantially from one state to another such that use
of a single state’s CEM for other states could lead to an erroneous conclusion regarding DSL versus USL safety impacts. However, by
looking at results for multiple states, this study shows how that problem is avoided. Two considerations for future applications of the
empirical Bayes methodology that proved helpful for this investigation are noted. First, it is feasible to use a site as its own control group
in a before/after study if the geographical variation from state to state is greater than the temporal variation from the before to the after
period. Such a situation arises not only in the case of statewide speed limit policies, but also for other transportation-related policies
enacted on a statewide basis. Second, results were consistent between conventional approaches and the modified empirical Bayes
approach. The latter technique showed the crash frequency increasing regardless of whether a state changed from DSL to USL, changed
from USL to DSL, maintained USL, or maintained DSL, leading one to conclude that speed limit policy has no consistent impact on
safety. This conclusion is similar to that reached when using conventional methods, which had shown that in each state, speed limit policy
共DSL versus USL兲 usually had no significant impact on crash rate 共crash frequency divided by vehicle miles traveled兲.
DOI: 10.1061/共ASCE兲0733-947X共2006兲132:1共19兲
CE Database subject headings: Traffic speed; Traffic safety; Trucks; Vehicles; Rural areas; Highways.

Introduction A policy question that has received significant attention is


whether passenger cars and trucks should have a uniform speed
Almost a decade has elapsed since the 1995 repeal of the National limit 共USL兲 or a differential speed limit 共DSL兲. Given that ton-
Maximum Speed Limit, and almost two decades have elapsed miles of freight shipped by intercity truck more than doubled
since the 1987 Surface Transportation and Uniform Relocation from 1980 to 2000 and that an increasing proportion of freight is
Assistance Act allowed states to raise rural interstate speed limits
expected to be shipped by truck rather than by rail by 2025, the
above 89 km/h 共55 mi/h兲. Accordingly, individual states have
interaction of trucks and passenger cars within the context of
had many opportunities to experiment with different speed limits.
speed limit policy is increasingly relevant 共BTS 2000, 2003兲.
1
Although proponents of a differential limit note that heavy
Professor, Dept. of Civil Engineering, Univ. of Virginia, P.O. Box trucks require a larger stopping distance and stopping time than
400742, Charlottesville, VA 22904. E-mail: njg@virginia.edu passenger cars 共thus necessitating a lower truck limit兲, proponents
2
Senior Research Scientist, Virginia Transportation Research Council,
of a uniform limit contend that the higher height of truck drivers
530 Edgemont Rd., Charlottesville, VA 22903. E-mail: john.miller@
virginiadot.org compensates for this and further, such a limit reduces speed
3
Former Graduate Student, Dept. of Civil Engineering, Univ. of variance, thereby reducing crash types susceptible to speed varia-
Virginia, P.O. Box 400742, Charlottesville, VA 22904. E-mail: xin.sun@ tion such as rear-end collisions. The disparate findings from
jacobs.com the literature reflect this disagreement between proponents and
4
Former Graduate Student, Dept. of Civil Engineering, Univ. of opponents. Some studies have shown no substantial difference
Virginia, P.O. Box 400742, Charlottesville, VA 22904. E-mail: bo.yuan@ between uniform and differential speed limits: examples include
psivanness.com no crash differences between USL and DSL 共Idaho Transportation
Note. Discussion open until June 1, 2006. Separate discussions must
Department 2000兲, the finding that a posted differential limit was
be submitted for individual papers. To extend the closing date by one
month, a written request must be filed with the ASCE Managing Editor.
not related to the occurrence of truck crashes 共Hall and Dickinson
The manuscript for this paper was submitted for review and possible 1974兲, no significant differences between speed variances at DSL
publication on June 2, 2004; approved on March 18, 2005. This paper is and USL sites 共Harkey and Mera 1994兲, and no significant
part of the Journal of Transportation Engineering, Vol. 132, No. 1, differences in mean and 85th percentile speeds 共Freedman and
January 1, 2006. ©ASCE, ISSN 0733-947X/2006/1-19–29/$25.00. Williams 1992兲. Other studies, however, have shown a negative

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 19


impact of differential limits, such as potentially higher crash estimation model rather than the control group, but this option is
rates for DSL facilities 共Garber and Gadiraju 1989, 1991a兲, a high not feasible for the state-by-state comparison of rural speed limit
speed differential increasing severity for rear-end collisions policy, since limits usually apply statewide.兲
between cars and trucks 共Council et al. 1998兲, and the suggestion On the other hand, if one must compare crash frequencies
that DSL might have increased rear-end crash rates based on a among different states, it can be said that rural interstate highways
comparison of adjacent states 共Garber and Gadiraju 1991b兲. can facilitate the creation of a control group more than other
The accepted methodology for evaluating transportation safety types of facilities. Operational and geometric characteristics for
impacts has also changed. Historically, the conventional approach interstate facilities are standardized to a greater degree than
has been to compare changes in crash rates, defined as the number they are for arterial and collector facilities, and rural interstate
of crashes per million vehicle miles traveled 共VMT兲 on a given highways tend to be less affected by traffic congestion than by
segment of highway. Within the past decade, however, a new urban facilities. Thus if a control group across states will succeed,
conceptual approach, the empirical Bayes methodology which it is likely to involve rural interstate highways. 共Still, driver char-
compares the number of expected crashes and the number of acteristics and other factors affecting the occurrence of crashes
observed crashes, has been refined such that it is increasingly may differ from state to state.兲
being viewed as a standard technique for assessing safety impacts
共Hauer 1997兲. In fact, the literature suggests four benefits of the
empirical Bayes approach relative to crash rate-based methods: Problem Statement
• The correct mathematical distribution for crashes is used.
Instead of presuming normality as might be the case with This paper investigates the safety impacts of differential speed
analysis of variance, the negative binomial distribution is a limits on rural interstate highways through accomplishing two
better descriptor of crash frequency 共Hauer 1997; Persaud objectives.
et al. 1997; Hauer et al. 1988; Hauer and Persaud 1988兲. 1. To determine the impact on the number of crashes of using a
• It is not assumed that other conditions remain constant. The differential versus a uniform speed limit; and
estimated number of crashes at a particular site usually 2. To demonstrate the modifications necessary to the empirical
does not remain constant from 1 year to the next because of Bayes approach in order to make it suitable for policy ques-
variation in traffic volume, traffic flow characteristics, weather tions where spatial control groups within the same state are
not feasible.
conditions, driver attitudes, and a host of variables beyond the
researcher’s control. While conventional methods strive to
control for these factors by judiciously selecting sites common
in all characteristics except that being studied, the empirical Methodology
Bayes technique ideally uses trend data to compensate for this
unwanted experimental variation 共Persaud et al. 2000兲. Conceptual Overview of the Empirical Bayes Approach
• A proportional relationship between crashes and average daily
Crash and volume data were synthesized from 17 different
traffic 共ADT兲 is not assumed. Instead, the number of crashes,
states to obtain a representative sample of interstate highways.
or crash frequency, is used to reflect highway safety, given
Unfortunately, useable speed data were not available except in the
that empirical studies have suggested that a percentage change
limited cases of traffic monitoring sites; details of the full data
in ADT will not necessarily have the same percentage change
collection process are not repeated here but are available else-
in crashes, even if all other factors were held constant. where 共Garber et al. 2003a,b兲. The six states with suitable data
• Crash estimation models are used to account for the fact comprise four policy groups as shown in Table 1.
that the observed crash frequency is just a point observation 1. Maintain a uniform limit for cars and trucks 共Arizona and
from some underlying distribution. In sum, the actual North Carolina兲;
number of crashes observed may not be an unbiased estimator 2. Maintain a differential limit for cars and trucks
of the expected number of crashes. Thus in lieu of this point 共Washington兲;
estimate, a crash estimation model 共CEM兲 is employed 3. Change from a uniform to a differential limit 共Arkansas and
to statistically predict the best estimator of crash frequency Idaho兲; and
共Persaud 1991; Bonneson and McCoy 1993; Hauer 1997; 4. Change from a differential to a uniform limit 共Virginia兲.
Persaud and Lord 2000兲. The premise behind the empirical Bayes approach is that one
Yet the empirical Bayes approach is not a panacea, given devises a crash estimation model that successfully predicts the
that its validity depends on the creation of appropriate crash crashes that would have occurred had no change in speed limit
estimation models, which in turn requires the selection of a policy been made. One then compares the number of these
representative control group. As presented, the control groups “expected after period” crashes to the number of actual crashes to
within the empirical Bayes technique addresses spatial rather estimate the impact of the change in speed limit policy. Four
than temporal variation. For example, to compute the benefits of discrete steps are thus required for the empirical Bayes approach:
reducing speed limits on local streets in a single town, one could 1. Create and assess a crash estimation model for each state
collect crash data from streets where such limits had been reduced that predicts the number of crashes at each site within that
and compare these data to other streets where such limits had not state as a function of related independent variables. In this
been reduced. The challenge within the context of differential case, those variables were traffic volume and segment
speed limits, however, is that such limits are usually enacted length. The notation for the crash estimation model at a given
statewide, such that it is impossible to select “adjacent” roadway site i for a given year y is E共mi,y兲. This step also entails
segments within the same state where both DSL and USL are verification that the crash estimation model is based on the
represented. 共It should be clarified that the empirical Bayes correct mathematical distribution and provides a reasonable
approach does address temporal correlation through the crash goodness of fit.

20 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006


Table 1. Rural Interstate Speed Limits: 1991–2000
State Speed limit policy Number of sites Category for this study
a
Arizona 121 km/h 共75 mi/h兲 278 Maintained uniform limit

North Carolina 105 km/h 共65 mi/h兲 before 1996 26 Maintained uniform limit
113 km/h 共70 mi/h兲 after 1996

Washingtonb 105/97 km/h 共65/60 mi/h兲 9 Maintained differential limit

Arkansas From: 105 km/h 共65 mi/h兲 10 Changed from uniform to differential limit
To: 113/105 km/h 共70/65 mi/h兲 1996

Idaho From: 105 km/h 共65 mi/h兲 32 Changed from uniform to differential limit
To: 121 km/h 共75 mi/h兲 1996
To: 121/105 km/h 共75/65 mi/h兲 1998

Virginia From: 105/89 km/h 共65/55 mi/h兲 266 Changed from differential to uniform limit
To: 105 km/h 共65 mi/h兲, 1994
a
Prior to December 1995, Arizona’s uniform limit was 105 km/h 共65 mi/h兲. This limit was raised by route to 121 km/h 共75 mi/h兲 between December 1995
and the summer of 1996.
b
Washington State limits shown refer to nine sections on I-90 which comprise the entire Washington data set used for this study. During inclement weather,
speed limits on those sections drop to 25 mi/h.

2. Determine the number of expected after period crashes at the reference group will capture that trend. The disadvantage is
each site during the after period. The number of crashes at that the chosen state would need to be similar to Virginia in
each site i for each year y is mi,y, and the summation of these terms of geometric, traffic, and population characteristics.
expected after period crashes, had no change in speed limit • Use data from the before period but for the same state. The
policy occurred, is ␲. advantages and disadvantages are the reverse of the previous
3. Compute the sum of the actual crashes that did occur at option: in this case, one is usually guaranteed of having the
each site i and for each year y共Ki,y兲 where the summation is right geometric characteristics, but one misses temporal trends
designated as ␭. that are present in the after years but not present in the before
4. Compute the ratio of the total actual crashes 共␭兲 to the total years.
expected after period crashes 共␲兲, and determine whether the The investigators chose the latter course of action, deeming
resultant index of effectiveness 共␪, which is numerically the disparity between states as being greater than the disparity
similar but not identical to ␭ / ␲兲 is significantly different between time periods. To illustrate this disparity, compare two
from unity by using the appropriate confidence intervals. crash models from the before periods for two different states from
By performing this four step approach for each state and for the same time period of 1991–1993: Virginia and Washington
various crash types within each state, one can assess whether State, both of which maintained a differential speed limit at that
there is a correlation between safety impacts and speed limit time.
policy. As a check on the veracity of this approach, one can also
use it for states where there was no speed limit change 共Arizona, Virginia: E共m兲 = 0.022共length兲0.622共ADT兲0.548
North Carolina, and Washington兲 to assess the impact of the
decision not to change the speed limit policy.
Washington: E共m兲 = 0.531共length兲0.440共ADT兲0.340

Step 1: Develop the Crash Estimation Model


A perfect crash estimation model determines the number
of crashes by site and by year, accounting for all confounding
factors such as operational changes 共e.g., volume growth,
enforcement modifications, and the installation of safety service
patrols兲, geometric changes 共e.g., work zones, median barriers,
or new interchange construction兲, and driver changes pertaining
to behavior, licensing, and vehicle maintenance. Unfortunately,
due to limited data available for this study, researchers had two
options for creating a CEM:
• Use data from the same time period but from a different state.
For Virginia one could select crash data for the period 1995–
1999 from a different state with comparable characteristics. Fig. 1. Comparison of crash estimation models for Virginia
The advantage is that if there is some national temporal trend, and Washington State 共1991–1993兲

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 21


Table 2. Comparison of Crash Estimation Models for the Before Periods correlation. This expression was ultimately not used because
for Other States reliable estimates could not be obtained for the “after years” for
State Crash estimation models the states that changed from uniform to differential or vice versa,
and the disparity in CEMs such as that shown in Table 1 hindered
Policy: Change from uniform to differential
“borrowing” ␣y coefficients from one state for use in another.
Arkansas E共m兲 = 0.00267共length兲1.0共ADT兲0.714 That is, for a state such as Virginia which changed from differen-
Idaho E共m兲 = 0.00258共length兲0.805共ADT兲0.745 tial to uniform in 1994, calibration of this model would indeed
Policy: Maintain uniform limit yield ␣y coefficients for the before period 共␣1991, ␣1992, and ␣1993兲.
Arizona E共m兲 = 0.415共length兲1.087共ADT兲0.127 However, since the entire Commonwealth of Virginia changed
North Carolina E共m兲 = 0.001共length兲0.772共ADT兲0.827 from DSL to USL, there were no sites that could be used to
estimate the coefficients for the after years 共e.g., ␣1995, ␣1996,
␣1997, and ␣1999兲 for DSL in Virginia. The only way to determine
Visual inspection of the plots of the crash estimation models, ␣1995, ␣1996, ␣1997, and ␣1999 for a DSL case would be to obtain
shown in Fig. 1 for a fixed length of 5 miles and a range of ADTs, them from a DSL state. Although one might then be tempted
confirms that different numbers of crashes at two sites, one in to borrow ␣y coefficients from another state and apply them
Virginia and one in Washington, with the same volume could to Virginia, Table 2 suggests that borrowing coefficients from
be predicted. For example, suppose a 5 mile site registered an one state to apply to another is problematic, even for the same
ADT of 10,000: the CEM for Virginia would have predicted time period.
approximately nine crashes whereas the CEM for Washington To ascertain whether the model was appropriate, the method
would have predicted almost 25 crashes. of cumulative residuals was used 共Persaud and Lord 2000兲.
Generally, the variation between states in terms of the crash 共The cumulative residual is the difference between the modeled
estimation models was found to be relatively large, even among number of crashes and the actual number of crashes for a
states with similar speed limit policies. Table 2 clearly indicates particular value of the independent variables in the CEM.兲 If the
that the Arizona or North Carolina models could not have cumulative residuals oscillate around 0 within the range of the
served as the predictors for the expected after period crashes for two plots of two standard deviations, a good quality of fit is
Arkansas and Idaho during the after period. For the crash estima- reflected 共Lord, personal communication, 2002兲. Figs. 2 and 3, for
tion models for specific types of crashes 共e.g., for truck involved example, reflect a good fit for Virginia CEMs based on section
crashes only兲 comparable dissimilarities between the states were length and AADT.
observed, even with the same speed limit policy. Finally, although recent studies have indicated that the
Thus, for this specific application, capture of nontreatment negative binomial distribution is a better descriptor for crash
sources of variation was best accomplished by using CEMs frequencies between sites, the investigators confirmed this by
developed within the same state. As this is a significant departure comparing the actual crash frequencies for Arizona, Idaho,
from the empirical Bayes formulation given in the literature, the North Carolina, and Virginia data with theoretical frequencies
nomenclature modified empirical Bayes is used, to credit others using the Poisson and negative binomial distributions. It was
with developing the empirical Bayes framework but to acknowl- found that the negative binomial distribution is a valid descriptor
edge that the selection of comparison sites herein deviates from of these data sets.
that original methodology.
The crash estimation model took the form shown in Eq. 共1兲,
Step 2: Estimate the Number of Expected
where the expected mean value of crashes E共mi,y兲 for a roadway
After Period Crashes
segment during a given year was the function of the length and
ADT. As recommended in the literature, the maximum likelihood After the CEM was developed, the expected crash frequency
technique was used to estimate the parameters since the crash m1 , m2,. . . , my for the treated segments for the after years
distribution was thought to follow a negative binomial distribu- were determined. To accomplish this, the following steps were
tion 共Hauer 1997; Persaud and Lord 2000兲. Note parameters ␤1 undertaken in sequence:
and ␤2 do not imply a proportionate effect of ADT and length • The CEM was applied to the “before” years data to obtain
unless equal to unity, since others have suggested that such an E共mi,y兲 as shown in Eq. 共1兲.
effect cannot be assumed 共Brown and Tarko 1999; Mountain et al. • The ratio Ci,y was then computed for each of these before
1996兲. years from y = 1 , 2 , . . . , Y. Thus Ci,y is the ratio of the current
E共mi,y兲 to the first year E共mi,1兲.
E共mi,y兲 = ␣共length兲␤1共ADT兲␤2 共1兲
Estimation of the parameters was done with the GENSTAT
software package 共release 4.2.1兲. The procedure also gives the E共mi,y兲
Ci,y = 共2兲
variance, VAR共mi,y兲 of the expected crash frequency for each site, E共mi,1兲
enabling the determination of an aggregation parameter 共k兲 which
denotes the extent to which the crash distribution is negative
binomial 共low value of k兲 or Poisson 共infinite value of k兲 共Hauer Thus for Virginia, Ci,1991; Ci,1992; and Ci,1993 were computed
and Persaud 1988; Hauer 1992兲. Software for estimating k devel- since Virginia’s before data were the 3 years 1991–1993. As
oped by Persaud and Lord was used for this effort 共Persaud and written, Eq. 共2兲 shows how the temporally variant aspects of a
Lord 2000兲. given site 共e.g., ADT, weather, truck percentages兲 influence its
Eq. 共1兲 was not chosen arbitrarily. The investigators consid- crash risk. By using year 1 in the denominator, one can easily
ered a variety of formulations including an accounting for see how crash risk changes relative to year 1 for a given site.
year-by-year changes such that ␣ in Eq. 共1兲 was replaced by ␣y, However, one could in fact use any year as the base and still
which, had it proven feasible, could have addressed temporal obtain the same measure of risk from the CEM 共Hauer 1997兲

22 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006


Fig. 2. Plot of goodness of fit for the crash estimation model versus ADT 共Virginia data兲

. mi,y = Ci,ymi,1 共5兲


• The expected crash frequency mi,y and its variance VAR共mi,y兲
for years of y = 1 , 2 , . . . , Y were then calculated from
Eqs. 共3兲–共6兲 共Hauer 1997兲. VAR共mi,y兲 = 共Ci,y兲2VAR共mi,1兲 共6兲
With the expected crash frequency m1 , m2,. . . , my estimated for
兺y=1 Ki,y
Y
k+ before period for each treated site from Eq. 共5兲, one can then
mi,1 = 共3兲 compute the “expected after period crash frequency”
k

Y
+ C mY+1 , mY+2,. . . , mY+Z for the after period as if there had been no
E共mi,1兲 y=1 i,y
such treatment using the following sequential steps.
where Ki,y⫽actual of crashes that occurred at site i in year y and • The crash estimation model from Eq. 共1兲 was applied to the
k⫽aggregation parameter for the negative binomial distribution after years data in order to obtain E共mi,y兲 for from y = Y + 1 to
y = Y + Z.
兺y=1 Ki,y
Y
k+ • The ratios Ci,y were computed using Eq. 共2兲 for the years from
mi,1

冉 冊
VAR共mi,1兲 = 2 = 共4兲 y = Y + 1 to y = Y + Z.
k k • The values for mi,y and VAR共mi,y兲 were computed using
兺y=1 Ci,y 兺y=1 Ci,y
Y Y
+ +
E共mi,1兲 E共mi,1兲 Eqs. 共5兲 and 共6兲 for the years from y = Y + 1 to y = Y + Z. These

Fig. 3. Plot of goodness of fit for the crash estimation model versus length 共Virginia data兲

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 23


Table 3. Crash Estimation Model Parameters for Virginia Data
State Parameter Total Fatal Rear-end Truck total Truck fatal
Virginia 共DSL to USL兲 k 5.9 5.58 1.71 3.99 15.0
␣ 0.02243 0.00119 0.000325 0.000316 0.000017
␤1 0.622 0.842 0.536 0.819 1.226
␤2 0.548 0.398 0.829 0.788 0.639
␪ 1.15 1.06 1.16 1.25 1.13

Arkansas 共USL to DSL兲 k 19.03 25.0 14.86 18.2


␣ 0.00267 0.0598 0.000000016 0.0239
␤1 1.0 1.0 1.0 1.0
␤2 0.714 0.00355 1.774 0.401
␪ 1.07 1.61 0.94 1.31

mi,y are the crashes that would have occurred had no treatment
been made. For the case of Virginia, whose after period was 冉 Var共␭兲 Var共␲兲
␭2
+
␲2

冉 冊
1995, 1996, 1997, and 1999, therefore, then the outcome of Var共␪兲 = ␪2 共11兲
this step would be, for each site, mi,1995, mi,1996, mi,1997, and Var共␲兲 2
1+
mi,1999. ␲2

Step 3: Compute the Sum of the Actual Crashes


that Did Occur confidence bounds共␪兲 = ␪ ± 2冑Var共␪兲 共12兲
The effect of the treatment is quantified by comparing the The confidence bounds are used to determine whether the
expected after period crashes 共shown above as mi,y for y ranging value for ␪ shows a statistically significant safety impact. If
from Y + 1 to Y + Z兲 to the actual crashes in the after period the confidence bounds for ␪ contain 1, then the safety impact
共shown above as Ki,y with y ranging from Y + 1 to Y + Z兲. The computed by Eq. 共11兲 are not significant; thus one cannot say that
expected after period crashes at each site i during the after period the treatment had a measurable effect.
are denoted as ␲i while the actual crashes at each site during the
after period are denoted as ␭i. Eqs. 共7兲 and 共8兲 are used to sum
the crashes from the individual sites as ␲ and ␭. Results and Discussion
␲= 兺 ␲i The empirical Bayes approach is best suited for a before/after
共expected after period crashes had there been scenario where some treatment is implemented, such as a change
in speed limit policy. Thus the results are presented first for policy
no speed limit change兲 共7兲
groups 3 and 4 where states changed from either a differential
limit to a uniform limit or from a uniform to a differential limit.
␭= 兺 ␭i
共actual crashes given that the speed limit did change兲 共8兲 Virginia Results „Differential to Uniform Speed Limit…
For the 266 sites which comprised the Virginia study data,
Step 4: Determine if the Ratio of Actual to Expected the length of the sites ranged from 1.05 to 14.25 mi in length.
After Period Crashes is Statistically Significant Calibration of the parameters for the crash estimation model
of the form shown in Eq. 共1兲 as E共m兲 = ␣共length兲␤1共ADT兲␤2
Eq. 共9兲 computes this ratio, also known as the index of effective-
are shown in Table 3 for the Virginia data. The fact that ␤2 is less
ness 共␪兲, such that a value of ␪ less than 1.0 indicates that the
than 1.0 for all crash types means that the model presumes
speed limit change improved safety, and the unbiased estimator of
that ADT does not have an equal proportional effect on crashes.
this index is Eq. 共10兲 共Hauer 1997兲. However, this particular
Instead, a certain percentage increase in ADT will result in a
study found the difference between Eqs. 共9兲 and 共10兲 to be quite
small. Eqs. 共11兲 and 共12兲 show the variance and confidence small percentage increase in the number of crashes. In short, an
bounds, respectively, for ␪. 共Note that ␪ is the “actual crashes” increase in ADT should yield a lower crash rate according to the
divided by expected after period crashes.兲 model. Similarly, the ␤1 values below 1.0 mean that an increase
in section length, according to the model, will increase crashes by
␭ a smaller percentage. 共For all states, usually the ␤1 exponent was
␪= 共9兲 also almost always less than 1.0, meaning that as the section

length increases, the number of crashes forecasted by the crash



冉冊 estimation model will correspond to a lower crash rate. Given
that this exponent was usually significant, its nonunity value is

冉 冊
␪= 共10兲 not attributable to random variation. Instead, that exponent is a
Var共␲兲 reflection of the differences in the section length upon which the
1+
␲2 crash data are based.兲

24 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006


Table 4. Total Crashes for Virginia
␭ ␭ ␲ ␲ VAR共␲兲 VAR共␲兲 ␪ ␪ Var共␪兲 Var共␪兲
Year 共by year兲 共cumulative兲 共by year兲 共cumulative兲 共by year兲 共cumulative兲 共by year兲 共cumulative兲 共by year兲 共cumulative兲
1995 3,608 3,608 3,223 3,223 986 986 1.119 1.119 0.000466 0.000466
1996 3,964 7,572 3,296 6,519 1,031 2,017 1.202 1.161 0.000502 0.000242
1997 3,735 11,307 3,383 9,903 1,088 3,105 1.104 1.142 0.000442 0.000156
1999 4,070 15,377 3,463 13,366 1,142 4,247 1.175 1.150 0.000471 0.0001175
Average 3,844 3,341 4,247 1.150

The applications of Eqs. 共5兲–共8兲 are reflected in Table 4. The showed ␪ less than 1.0; further, note that it was only rear-end
application of Eq. 共7兲 results in the actual crash data for each year crashes that had a ␤2 ADT coefficient that was greater than 1.0.
being shown in the ␭ column with the cumulative values of ␭
shown in the ␭ cumulative column. Similarly, the application of
Crashes from the States of Idaho, Arizona,
Eq. 共8兲 shows the predicted expected after period crash data for
North Carolina, and Washington
each year during the after period listed in the fourth column 共␲兲
with the cumulative values of these expected after period crashes Table 5 shows the values of ␪ that were obtained when evaluating
shown in the fifth column 共␲ cumulative兲. The variance of ␲ was the effect of maintaining the same speed limit in four other states.
calculated for each year during the after period and listed in the For comparison purposes, the values from Virginia and Arkansas
sixth column 关VAR共␲兲兴, with the cumulative values of VAR共␲兲 are also shown. In most cases, Table 5 shows that ␪ was greater
listed in the next column 关VAR共␲兲 cumulative兴. The evaluation than 1.0. Examining these ␪ values for individual states, it
was investigated by using Eq. 共10兲 to obtain the ratio 共␪兲 of the seems that although the ratio for total crashes in Virginia, which
actual after crashes to the expected after period crashes. The ␪ by changed from DSL to USL, is higher than that for Arkansas
year column indicates the value obtained from that year alone, 共which changed from USL to DSL兲, it is lower than that for Idaho
whereas the ␪ cumulative column indicates the ratio that would 共which also changed from USL to DSL兲. In Virginia the most
result from using all crashes from 1995 through the given year. important factor is the total fatal crash ratio which is 1.06, while
Variances for the ratio ␪ are listed to the right of Table 4 as an that for Arkansas which changed from USL to DSL was 1.61.
application of Eq. 共11兲. In fact all the ratios for Idaho which changed from USL to DSL
Table 4 shows that the ratio ␪ for total crashes in Virginia was were much higher than those for Virginia. This suggests that
larger than 1.0. Further, application of Eq. 共12兲 with Var共␪兲 although there was an overall trend for an increase in crashes,
= 0.0001175 yields empirical confidence bounds for ␪ very close the percentage increases, particularly for fatal crashes, tend to be
to 1.15; these bounds certainly do not include 1.0. Thus according higher in states that changed from USL to DSL.
to the empirical Bayes technique, since the ratio of the actual after On the other hand, the data from Table 5 are not consistent.
crashes 共␭兲 to the expected after period crashes 共␲兲 is greater than For total crashes, it is interesting to note that ␪ remained approxi-
1.0, then the treatment 共a change from a differential limit to a mately 1.0 for the state that maintained DSL 共Washington兲 while
uniform limit兲 resulted in an increase in the number of crashes. In it was greater than 1.0 for states that maintained USL 共Arizona
fact, for all Virginia crash types, ␪ is greater than 1.0. If one had and North Carolina兲. At a first glance, therefore, the reader might
restricted the analysis to Virginia alone, then the interpretation of draw the interpretation that maintaining DSL caused no change in
these results would have been that the change to a uniform speed crashes while maintaining USL caused an increase in the number
limit increased the number of crashes; however, as is explained in of crashes. Examination of the fatal crashes, however, shows
the following sections, data from other states do not support this inconsistency in those states that maintained USL where fatal
interpretation. crash data were available: ␪ was greater than 1.0 for Arizona
but less than 1.0 for North Carolina. Again, comparing the
impacts for Arkansas and Virginia—two states with diametrically
Arkansas Results „Uniform to Differential Speed Limit…
opposite policies—showed that such policies did not have
Arkansas is of special interest because its speed limit policy opposite impacts on crashes. For example, for total crashes
is the reverse of that just discussed: Arkansas changed from a involving trucks, the two states had similar values for ␪ 共1.31 and
uniform to a differential speed limit. The Arkansas data are from 1.25兲 despite the fact that the former changed from USL to DSL
10 interstate sections in Arkansas, each of 10 miles in length. and the latter changed from DSL to USL.
As shown in Table 3, the ␤1 parameter was assumed to be 1.0 Interpretation of the effectiveness index ␪ as described here
since all sections were of the same length and the ␤2 parameter should be tempered by consideration of the confidence bounds
was found to be less than 1.0 for all crash types except that of from Eq. 共12兲: is there a significant change in the number of
rear-end crashes. The 1.774 value for the ␤2 coefficient means crashes between the before and after periods, or is the difference
that, according to the model, a certain increase in ADT would from unity merely ascribed to random variation?
increase rear-end crashes by a larger proportion: in short, at a
given site, increasing the ADT will increase the crash rate.
Are the Safety Impacts Statistically Significant?
The last row of Table 3 summarizes the resultant safety
impacts. Except for rear-end crashes, all crash types for Arkansas Table 6 presents confidence intervals for the expected percentage
showed that ␪ was significantly greater than 1.0, meaning that the crash increase for each type of crash and for each state according
shift 共from uniform to differential兲 increased the number of to the empirical Bayes method. For example, consider the cat-
crashes. Of interest is the fact that rear-end crashes were the only egory of “total crash with truck involved” for Arkansas which
crash type that did not follow this trend, since that crash type changed from uniform to differential and Virginia which changed

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 25


Table 5. Impact of Speed Limit Changes According to the Empirical Bayes Formulation
State Crash type Ratio ␪ ␤2 Inference based on study of this state in isolationa
Policy Group 4: Changed from differential to uniform
Virginia Total crash 1.15 0.548 According to Virginia data, the actual number
Total crash with truck involved 1.25 0.788 of crashes during the after period was higher
Rear-end crash 1.16 0.829 than what would have been expected without
the change to a uniform speed limit.
Fatal crash 1.06 0.398
Fatal crash with truck involved 1.13 0.639
Average 1.15
Policy Group 3: Changed from uniform to differential
Arkansas Total crash 1.07 0.714 According to Arkansas data, the actual number
Total crash with truck involved 1.31 0.401 of crashes during the after period was higher
Rear-end crash 0.94 1.774 than what would have been expected without
the change to a differential speed limit, except
Fatal crash 1.61 0.00355
for rear end crashes.
Average 1.23

Idaho Total crash 1.29 0.745 According to Idaho data, the actual number
Total crash with truck involved 2.46 0.717 of crashes during the after period was higher
Rear-end crash 1.62 1.717 than what would have been expected without
the change to a differential speed limit.
Rear-end crash with truck involved 2.36 1.698
Average 1.93
Policy Group 1: Maintained a uniform limit
Arizona Total crash 1.26 0.127 According to Arizona data, the actual number
Total crash with truck involved 1.16 0.208 of crashes during the after period was higher
Rear-end crash 1.2 0.757 than what would have been expected even
though the uniform speed limit was maintained.
Rear-end crash with truck involved 1.07 0.537
Fatal crash 1.33 0.120
Fatal crash with truck involved 1.63 0.180
Average 1.28

North Carolinab Total crash 1.26 0.827 According to North Carolina data, the actual number
Total crash with truck involved 0.91 1.704 of total crashes during the after period was higher
Rear-end crash 1.002 1.619 than what would have been expected though a uniform
speed limit was maintained, except for some specific
Rear-end crash with truck involved 0.97 1.834
types of crashes where the opposite was observed to
Fatal crash 0.74 2.878 increase total crashes and to decrease fatal crashes.
Average 0.98
Policy Group 2: Maintained a differential limit
Washington Total crash 0.99 0.340 According to Washington data, the decision to maintain
a differential speed limit had no effect on crashes.
a
The inferences shown are the findings that would be drawn if each state were examined in isolation.
b
North Carolina maintained their uniform limit but also raised this limit for both passenger cars and trucks.

from differential to uniform. Table 5 shows ␪ values of 1.31 and limit in Virginia did not affect “truck-involved fatal crashes”
1.25 for Arkansas and Virginia, respectively, suggesting that there 共since that confidence interval runs from −12.5 to + 38.4% and
was a 31% increase in Arkansas and a 25% increase in Virginia thus includes zero兲 but that it did cause total crashes to signifi-
as a result of the changes in the speed limit policies. However, cantly increase 共since the corresponding confidence interval is
these percentage increases are not perfect, given the variability 12.9–17.2% and thus does not include zero兲. Similarly, the change
that can occur in crashes. Thus Table 6 is used to suggest that in Arkansas increased total crashes significantly. Interestingly,
the confidence interval for Arkansas is between 18.9 and 42.8%, confidence intervals associated with the empirical Bayes method
and between 20.0 and 29.8% for Virginia. As can be inferred from in Table 6 showed either increases in crashes or no change in
examination of crash types presented in Table 6 later in this paper, crashes, for all states and for all categories of crashes regardless
larger sample sizes tended to lead to narrower confidence inter- of the speed policy change, with only one exception: fatal crashes
vals 共e.g., within a state, the confidence interval for the many in North Carolina, a state which maintained a uniform speed
rear-end crashes was tighter than the corresponding interval for limit.
the few fatal crashes兲. The impact of differential speed limits as studied by conven-
Generally, the standard interpretation of confidence intervals tional methods has already been reported elsewhere where it was
is to declare the change insignificant if the confidence interval found that differential speed limits had no significant safety
includes zero. Using the Virginia example, therefore, one would impact relative to uniform speed limits. It is thus possible to
infer that according to this application of the empirical Bayes compare the results of this modified empirical Bayes approach to
methodology, the change from a differential to a uniform speed conventional approaches. For example:

26 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006


Table 6. Crash Increases and Confidence Intervals According to the Empirical Bayes Formulation
Lower Upper
bound increase bound increase Statistical
State Crash type 共%兲 共%兲 interpretation
Policy Group 4: Changed from differential to uniform
Virginia Total crash 12.9 17.2 Increase
Total crash with truck involved 20.0 29.8 Increase
Rear-end crash 11.2 21.1 Increase
Fatal crash −7.3 19.9 No change
Fatal crash with truck involved −12.5 38.4 No change
Policy Group 3: Changed from uniform to differential
Arkansas Total crash 0.4 13.4 Increase
Total crash with truck involved 18.9 42.8 Increase
Rear-end crash −16.7 3.9 No change
Fatal crash 1.5 121.3 Increase

Idaho Total crash 13.2 46.7 Increase


Total crash with truck involved 68.6 224.9 Increase
Rear-end crash 30.6 94.8 Increase
Rear-end crash with truck involved −7.4 281.1 No change
Policy Group 1: Maintained a uniform limit
Arizona Total crash 24.2 28.6 Increase
Total crash with truck involved 12.1 20.7 Increase
Rear-end crash 14.8 25.3 Increase
Rear-end crash with truck involved −1.5 14.7 No change
Fatal crash 20.6 45.2 Increase
Fatal crash with truck involved 30.0 97.4 Increase

North Carolina Total crash 19.9 31.9 Increase


Total crash with truck involved −19.7 1.5 No change
Rear-end crash −12.6 13.1 No change
Rear-end crash with truck involved −23.4 18.6 No change
Fatal crash −50.1 −2.1 Decrease
Policy Group 2: Maintained a differential limit
Washington Total crash −6.6 5.0 No change

• Under the conventional approach, an analysis of variance with lowering of truck speed limits, as well as the aforementioned
30 cases 共based on five states and six crash types per state, work that saw no consistent change in crash rates despite changes
such that all four policy groups were represented兲 showed that in speed limits 共Garber et al., 2003a,b兲.
for 27 of the 30 cases, there was no significant difference in
crash rates 共Garber et al. 2003a,b兲. Such a finding suggests
Caveats to the Use of the Modified
speed limit policy had no consistent safety impact according to
Empirical Bayes Method
conventional methods; and
• In the modified empirical Bayes approach, Table 6 generally Three caveats are noted for future applications of the modified
showed a significant increase in the number of crashes regard- empirical Bayes method based on the results shown in Tables 5
less of speed limit policy, leading one to conclude that because and 6.
crashes increase regardless of policy, the policy itself has no First, interpretation of the results should be influenced by com-
consistent effect. prehensive study of the crash estimation model. The CEMs used
Therefore the conclusions drawn with conventional approaches in this study were appropriate based on goodness of fit tests for
and those drawn with the modified empirical Bayes approach are total Virginia crashes 共Figs. 2 and 3兲 and significance testing.
similar. In short, use of the modified empirical Bayes approach Interpretation, however, is contingent on several aspects:
leads to a consistent finding with conventional approaches, • The coefficients of the crash estimation model which suggest
provided one looks at the results across multiple states, not just crash frequency is more insensitive to ADT than might be the
a single state. The use of these multiple states allows one to case in reality. Table 5 shows the exponent ␤2 which reflects
intuitively assess the validity of the CEMs that were developed. the impact of a change in ADT. Because ADT increases over
Further, careful examination of Tables 1 and 6 might lead one time for most states, a ␤2 less than 1.0 means the expected
to infer that increasing limits of any type is responsible for an after period crashes predicted by the CEM correspond to a
increase in crash risk, as suggested by the states of Arizona, North lower crash frequency for a given ADT. Thus according to
Carolina, Arkansas, and Virginia. However, this inference would the model, a large increase in ADT should only have a very
be refuted by Idaho which saw total crashes increase despite a small increase in the number of crashes—hence the model is

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 27


Table 7. Distribution of Cases by Change from the before to the after multiple years of data and multiple sites, one can have some
Perioda confidence in the model, but still temporal changes in the after
Significant Significant year to alter this relationship are possible.兲
decrease No increase Third, the selection of sites should be unbiased. In traditional
Exponent in the number significant in the number before/after studies, such a decision could possibly subject the
for ADT of crashes change of crashes study to what Hauer describes as “regression-to-mean 共RTM兲
␤2 ⬎ 1.0 1 5 1 bias” or “selection bias.” As explained by Hauer, RTM arises
␤2 ⬍ 1.0 0 4 14 when “there is a link between the decision to treat an entity and
a
Each row in Table 7 共e.g., each state and crash type兲 represents a case. its accident history” 共Hauer 1997兲. If the basis of site selection
was those sites with the largest number of crashes in a single
year, then it is possible that subsequent crash reductions would
be erroneously attributed to the treatment when in fact such
reductions were truly the result of a random variation that would
insensitive to ADT. In reality, however, crash frequency is at have transpired even without the treatment. Two characteristics
least somewhat sensitive to ADT. In cases with a low value of of this study, however, likely eliminate the possibility of regres-
␤2 far from unity, as was the case with Arkansas fatal crashes, sion to the mean bias. First, the sites studied were generally
the model suggests that large increases in ADT should only not chosen by persons with an interest in testing the effects of
increase crashes by a disproportionately small amount, such USL and DSL; rather, the sites were those that had available data.
that the expected after period crash frequency for a given ADT Second, the sites did not show behavior expected in a study where
is extremely low. In those cases, it is virtually impossible for regression to the mean occurs: instead of seeing dramatic crash
any policy change to show ␪ less than 1.0 because radical reductions, states generally saw crash increases. Had the bias
changes in crash frequency for a given ADT are generally not been present, one would have expected to see crash reductions in
observed on the Interstate system. In fact, Table 7 suggests, lieu of increases.
but does not prove, that having a value of ␤2 greater than 1.0
decreases the likelihood that a significant increase will be
found, based on the 25 cases outlined in Table 6. Certainly
Conclusions
an area appropriate for future investigation is the temporal
1. The modified empirical Bayes approach showed no consis-
stability of the relationship between ADT and crashes, as was
tent safety impacts attributable to differential or uniform
suggested by a reviewer of this paper.
speed limit policy for rural interstate highways.
• The significance of the coefficients in the CEM. For almost all
2. Most of the crash estimation models developed by the
states and crash types, either section length, section ADT, or
investigators for the modified empirical Bayes approach,
both were significant determinants of the number of crashes in
in this study, were very insensitive to changes in ADT.
the CEM. The notable exception is Arkansas fatal crashes,
In most cases for this study, as time passed, the actual
where the crash estimation model for that particular case was
number of crashes for the after period was larger than the
probably spurious given the extraordinarily low value of the
predicted expected after period crashes. The study suggests
corresponding T-statistic for the model parameter associated
this is the result of the less than unity exponent associated
with volume. Conversely, CEMs may ideally include other
with Eq. 共1兲.
data elements that affect crash frequency, such as the number
3. In order to assist decision makers under conditions of sparse
of interchanges, in other applications if such data become data availability, it may be productive to analyze safety
available. impacts through both the traditional method of crash rates
• The confidence interval associated with ␪. Table 6 indicates and the newer method of the modified empirical Bayes
the extent to which the change was significant. In the case of approach when challenging conditions for developing
the Arkansas fatal crashes, this change is suggested as being crash estimation models exist. If the aggregate results from
almost but not quite insignificant, with the lower bound of the the two methods are the same, as was the case when
confidence interval approaching zero. comparing the conclusions herein from that of conventional-
Second, comparison groups are imperfect. Ideally, the com- based approaches with the same data set, one can present
parison 共control兲 group is selected from the same state at the same results using the method that is more familiar to the
time as the studied group. For example, in Virginia, it would have decision maker.
been ideal if after the statewide differential speed limit of the 4. The study demonstrates the potential to develop control
early 1990s was repealed, one section of Interstate 81 could have groups when spatial variation is larger than temporal varia-
been kept at a differential speed limit and another section of I-81 tion. The disparity of the crash estimation models in Table 2
could been changed to a differential limit. Thus, since the actual was one reason for using this modified empirical Bayes
Virginia reference group was an extrapolation of the temporal approach, and such situations arise when a statewide policy,
trend that occurred during the before period from 1991 to 1993, such as speed limit, inhibits the creation of adjacent control
later aberrations in this trend may not have been identified. In groups within the same geographical area. This situation
other words, if some significant change occurred in, say, 1995 is not limited to speed limits; there are other examples
that increased the crashes but had nothing to do with speed 共licensure laws, traffic control standards, and work zone
limits policy change, then unfortunately this change would not configurations兲 where because an entire state or multistate
be captured in the comparison group and therefore would not be region follows the same approach, one cannot have control
reflected in the models. 共A problem would arise if the relationship groups at the same time with different policies. For these
between ADT and crashes had drastically changed from the situations, a modified empirical Bayes approach as presented
before to the after period. When this relationship is based on here merits consideration.

28 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006


Acknowledgments Garber, N. J., and Gadiraju, R. 共1991a兲. Impact of differential speed limits
on highway speeds and accidents, Dept. of Civil Engineering,
Univ. of Virginia, Charlottesville, Va.
The writers gratefully acknowledge the many persons who
Garber, N. J., and Gadiraju, R. 共1991b兲. Impact of differential speed limits
provided data, insights, assistance, and other contributions
on highway speeds and accidents, AAA Foundation for Traffic
to this effort. Any person who has worked with state crash and Studies, Washington, D.C.
speed data will recognize that obtaining these data, especially Garber, N. J., Miller, J. S., Yuan, B., and Sun, X. 共2003a兲. “The safety
data from as far back as a dozen years ago, is no small undertak- impacts of differential speed limits.” Transportation Research Record.
ing. Assistance from K. Bonds 共Ark.兲, J. Williams 共Ariz.兲, 1830, Transportation Research Board, Washington, D.C., 56–62.
C. Haack and J. Benton 共Calif.兲, N. Brizzi 共Id.兲, L. Rust 共Ind.兲, Garber, N. J., Miller, J. S., Yuan, B., and Sun, X. 共2003b兲. The safety
W. Baldwin, G. Groat, M. Pawlovich, and T. Welch 共Iowa兲, impacts of differential speed limits on rural interstate highways,
J. Schaefer 共Mo.兲, P. Jomini and V. Gonzales 共Mont.兲, S. Cosper Federal Highway Administration, Washington, D.C.
and A. Wyatt 共N.C.兲, L. Evans, D. Gardner, J. Hausman, and Hall, J. W., and Dickinson, L. V. 共1974兲. An operational evaluation of
S. Shokouhi 共Ohio兲, S. Reed 共Ore.兲, R. Collins and M. Moore truck speeds on interstate highways, Dept. of Civil Engineering,
共Tex.兲, B. Siddiqui 共Utah兲, T. Rickman and K. Tanner 共Wash.兲, Univ. of Maryland.
J. Legg, R. Lewis, and R. Russell 共W. Va.兲 proved essential to this Harkey, D. L., and Mera, R. 共1994兲. Safety impacts of different speed
limits on cars and trucks, Federal Highway Administration,
study. Insights from D. Lord and E. Hauer also proved essential
Washington, D.C.
for application of the empirical Bayes method. Comments from Hauer, E. 共1992兲. “Empirical Bayes approach to the estimation of
D. Warren and R. Kremmes 共Federal Highway Administration兲 ‘unsafety’: The multivariate regression method.” Accid. Anal Prev.,
and F. Council 共N.C.兲 were also useful. Finally, the team also 24共5兲, 457–477.
could not have accomplished this work without the data reduction Hauer, E. 共1997兲. Observational before-after studies in road safety,
assistance of L. Woodson 共Va.兲. Inclusion of the above names and estimating the effect of highway and traffic engineering measures on
states does not imply, however, agreement with the results of this road safety, Dept. of Civil Engineering, Univ. of Toronto.
study. Hauer, E., Ng, J. C. N., and Lovel, J. 共1988兲. “Estimation of safety at
signalized intersections.” Transportation Research Record. 1185,
Transportation Research Board, Washington, D.C., 48–61.
Hauer, E., and Persaud, B. N. 共1988兲. “How to estimate the safety
References of rail-highway grade crossings and the safety effects of warning
devices.” Transportation Research Record. 1114, Transportation
Bonneson, J. A., and McCoy, P. T. 共1993兲. “Estimation of safety at Research Board, Washington, D.C., 131–140.
two-way stop-controlled intersections on rural highways.” Idaho Transportation Department Planning Division. 共2000兲. Evaluation
Transportation Research Record. 1401, Transportation Research of the impacts of reducing truck speeds on interstate highways
Board, Washington, D.C., 83–89. in Idaho—Phase III, Final Rep., National Institute for Advanced
Brown, H. C., and Tarko, A. P. 共1999兲. “The effects of access control on Transportation Technology, Univ. of Idaho.
safety on urban arterial streets.” Presented at the 78th Annual Meeting Mountain, L., Fawaz, B., and Jarrett, D. 共1996兲. “Accident prediction
of the Transportation Research Board, Transportation Research models for roads with minor junctions.” Accid. Anal Prev., 28共6兲,
Board, Washington, D.C. 695–707.
Bureau of Transportation Statistics 共BTS兲. 共2000兲. “The changing face Persaud, B. N. 共1991兲. “Estimating accident potential of Ontario road
of transportation.” BTS00-007, BTS, Washington, D.C. sections.” Transportation Research Record. 1327, Transportation
Bureau of Transportation Statistics 共BTS兲. 共2003兲. BTS 2003 pocket guide Research Board, Washington, D.C., 47–53.
to transportation highlights growth of air freight, BTS, Washington, Persaud, B., Hauer E., Retting, R., Vallurupalli, R., and Muxsi, K. 共1997兲.
D.C., 具http://www.bts.gov/PressReleases/2003/bts001_03.html典. “Crash reductions related to traffic signal removal in Philadelphia.”
Council, F. M., Duncan, C. S., and Khattak, A. J. 共1998兲. “Applying the Accid. Anal Prev., 29共6兲, 803–810.
ordered probit model to injury severity in truck-passenger car rear-end Persaud, B. N., and Lord, D. 共2000兲. “Accident prediction models
collisions.” Transportation Research Record. 1635, Transportation with and without trend: Application of the Generalized Estimating
Research Board, Washington, D.C., 63–71. Equations 共GEE兲 procedure.” Presented at the 79th Annual Meeting of
Freedman, M., and Williams, A. F. 共1992兲. “Speeds associated with 55 the Transportation Research Board, Transportation Research Board,
mph and 65 mph speed limits in northeastern states.” ITE J., 2共2兲, Washington, D.C.
17–21. Persaud, B. N., Retting, R. A., Gardner, P. E., and Lord, D. 共2000兲.
Garber, N. J., and Gadiraju, R. 共1989兲. Effect of truck strategies on traffic Crash reduction following installation of roundabouts in the United
flow and safety on multilane highways, AAA Foundation for Traffic States, Insurance Institute For Highway Safety, Ryerson Polytech
Safety. Univ., Toronto.

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2006 / 29

You might also like