Accident Analysis and Prevention 133 (2019) 105322

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

The more (sharp) curves, the lower the risk T

Rune Elvik
Institute of Transport Economics, Gaustadalleen 21, 0349, Oslo, Norway

A R T I C LE I N FO A B S T R A C T

Keywords: The risk of accident in horizontal curves is a complex function of at least the following characteristics of the
Horizontal curve curve: the radius of the curve; the length of the curve (and the resultant deflection angle); the presence of a spiral
Radius transition curve; the super-elevation of the curve; the distance to adjacent curves; and whether the curve is on a
Adjacent curve flat road, a straight gradient or a vertical curve. The interactions between these characteristics in determining
Risk
accident risk in horizontal curves is only beginning to be understood. This paper summarises the results of
Interactions
studies that have investigated the interaction between the radius of a horizontal curve and the distance to
adjacent curves. The shorter the mean distance between curves, the lower is the increase in risk for a given curve
radius. The sharper neighbouring curves are, the lower is the increase in risk for a given curve radius. Thus,
overall risk may not be higher on a road consisting mostly of sharp curves than on a road consisting mostly of
straight sections with a few curves located far apart from each other.

1. Introduction adjacent curves (Gooch et al., 2016, 2018); whether horizontal curves
are located on flat roads, straight gradients or vertical curves (Bauer
It has been known for a long time that the risk of accident is higher and Harwood, 2013; Saleem and Persaud, 2017); and the use of
in horizontal curves than on straight road sections. Perhaps the most warning signs or advisory speed limits in curves (e.g. Montella et al.,
widely studied characteristic of horizontal curves is their radius, i.e. 2015). No study has included all these factors. Hence, their contribu-
how sharp the curves are. Although studies have consistently found that tions to the safety of horizontal curves is not well known.
accident rate per million vehicle kilometres driven in curves increases The objective of this paper is to summarise evidence from studies of
as radius declines, there is large variation in estimates of the increase in the interaction between the distance between adjacent curves and the
risk. This is particularly the case when radius is less than 200 m. Thus, increase in risk in curves of a given radius. As noted above, the increase
Elvik (2013) noted that when relative risk is set to the value of 1.0 for a in risk in horizontal curves of a given radius varies substantially, and
curve radius of 1,000 m, it was found to vary between 2.6 and 8.2 in one of the factors associated with this variation is the distance between
curves with a radius of 100 m in six studies made in six different curves. The main research questions asked in the paper are:
countries. A review of recent North-American studies (Elvik, 2017)
found that the discrepancy in estimates of risk associated with sharp 1 Has an interaction between the number of horizontal curves on a
curves remains. Relative risk in curves with a radius of 150 m varied given length of road and the increase in risk for curves with a given
between 1.9 and 4.1 when the risk in curves with a radius of 1,200 m or radius been found consistently in studies examining this interaction?
more was set to 1.0. Clearly, risk in sharp curves is influenced not just 2 Does the interaction depend on the radius of a horizontal curve; i.e.
by their radius. is there a radius beyond which the interaction becomes negligible?
In addition to radius, accident rate in horizontal curves has been 3 What may explain the interaction between the number and sharp-
found to be influenced by: the presence of spiral transition curves ness of curves and the increase in risk associated with them?
(Zegeer et al., 1991; Tom, 1995); the length of curves (and the resulting
deflection angle) (e.g. Persaud et al., 2000; Saleem and Persaud, 2017; 2. Study retrieval
Bil et al., 2018); super-elevation in curves (e.g. Sakshaug, 1998;
Christensen and Ragnøy, 2006); road surface friction (Musey and Park, The studies identified by Elvik (2013, 2017) in previous reviews
2016); the distance to adjacent curves (Matthews and Barnes, 1988; were included. To identify new studies, searches were made of ISI Web
Eick and Vikane, 1992; Eriksen, 1993; Stigre, 1993; Hauer, 1999; of Science, Science Direct and Transportation Research Record online
Findley et al., 2012; Khan et al., 2013; Bil et al., 2018); the radius of using “horizontal curve” and “radius” and/or “accidents” or “crashes”

E-mail address: [email protected].

https://doi.org/10.1016/j.aap.2019.105322
Received 26 May 2019; Received in revised form 25 August 2019; Accepted 2 October 2019
0001-4575/ © 2019 Elsevier Ltd. All rights reserved.
R. Elvik Accident Analysis and Prevention 133 (2019) 105322

as search terms occurring in the title, abstract or key words of papers. The highest estimated accident rate (0.671) then gets the value of 5.32
Studies were included if they: (1) developed models or contained esti- (0.671/0.126).
mates of the association between the radius of horizontal curves and the
accident rate (or accident frequency) in the curves, and (2) developed 3.2. Findley et al. (2012)
models or contained data shedding light on the interaction between
radius and the distance to neighbouring curves in influencing the ac- The next study exploring how accident rate depends both on the
cident rate in curves with a given horizontal radius. distance between curves and their radius was reported by Findley et al.
It was not possible to statistically combine the results of different (2012). The study applied the CMF (crash modification function) de-
studies by means of standard techniques of meta-analysis. However, the veloped for the Highway Safety Manual:
results of the studies reviewed in the next section have been made
comparable by: (1) Converting estimates of accident frequency to ac-
CMF =
(1.55 ∙ Lc ) + ( )−
80.2
R
(0.012 ∙S )
cident rate (i.e. accidents per million vehicle kilometres); (2) (1.55 ∙ Lc ) (3)
Converting accident rates to accident modification factors by setting the
value of the lowest estimated accident rate in any study to 1.0 and In Eq. 3, Lc is the length of a curve in miles, R is the radius of the
expressing other accident rates as a multiple of this value. The results of curve in feet and S is an indicator for the presence of a spiral transition
the studies are then summarised in terms of functional relationships and curve, equal to 1 if there is a transition curve at both ends of a curve,
the shape of these relationships is compared graphically. 0.5 if there is a transition curve at one end only, and 0 if there is no
transition curve. When applying the equation in this paper, it was as-
sumed that there is no transition curve. It was further assumed that the
3. Estimation of risk – freqency of curves
length of a curve is equal to its radius, which implies that the deflection
angle is equal to one radian (57.3 degrees). Eq. 3 estimates a crash
A total of five functional relationships using distance between
modification function, i.e. a multiplicator showing how much higher
neighbouring curves as the independent variable and relative risk in
the accident rate per million vehicle miles is in a horizontal curve
curves with a given radius have been developed. This section explains
compared to a straight section.
how these relationships were estimated.
Findley et al. added a correction term to Eq. 3, defined as follows:

3.1. Matthews and Barnes (1988) – re-analysed by Hauer (1999) Correction= CMF e[B0 + (B1 ∙ D) + (B2 ∙ P ) + (B3 ∙ (D ∙ P ))] (4)

The correction term is an exponential function containing a constant

The oldest study that examined the interaction between the distance term (B0), a term for the distance to the distal curve(B1) (i.e. the
between curves and curve radius in influencing accident rate was re- neighbouring curve furthest away from a given curve), a term for the
ported by Matthews and Barnes (1988). Hauer (1999) re-analysed the distance to the proximal curve (B2) (i.e. the neighbouring curve closest
study, fitting the following models to describe its results: to a given curve) and a term (B3) for the interaction between distances
−6 2
R − 4.17 ∙ 10−3R) ∙ e (−(6.2 ∙ 10−4 − 1.2 ∙ 10−6R) ∙ (1200 − T )) to distal and proximal curves. Distances to distal and proximal curves
Accident rate = e (1.73 ∙ 10
were measured in miles. The results for distance to proximal curve of
(1)
0.3 miles and distances to distal curve between 0.3 and 2.1 miles have
Accident rate = e (1.73 ∙ 10
−6 2
R − 4.17 ∙ 10 −3
R) (2) been extracted. The multipliers (the exponential function in Eq. 4)
ranged from 1.267 for a distance of 0.3 miles to the distal curve to
R denotes the radius of a curve in metres and T denotes the length in 2.138 for a distance of 2.1 miles to the distal curve. Risk in a curve with
metres of the tangent (straight) section preceding a curve. Eq. 1 applies radius 1,200 m was used as reference (i.e. set to 1.0) when estimating
to curves with a radius less than 500 m and a tangent length less than accident modification factors.
1,200 m. Eq. 2 applies to curves with a radius of 500 m or more. No
correction for tangent length was applied to curves with radius larger 3.3. Khan et al. (2013)
than 500 m. Accident rate was stated as the number of accidents per
million vehicle kilometres of travel. Estimates of accident rate devel- Khan et al. (2013) estimated a set of models to predict accident rates
oped by means of Eqs. 1 and 2 have been tabulated in Table 1.
in curves. All models were negative binomial regression models of the
It is seen that the length of the tangent (straight) section ahead of a following basic form:
curve has a larger influence on accident rate the sharper the curve is.
The accident rate for a tangent length of 1,200 m and curve radius of Number of accidents = e (Coefficientsi ∙ Predictor variablesi) (5)
700 m (0.126, lowest rightmost cell of Table 1) is given the value of 1.0. The predictor variables included in the model referring to the lar-
gest accident data set (the ALL crash data set; N = 15,097 accidents)
Table 1
were:
Accident rates as a function of horizontal curve radius and length of straight
section (tangent) between curve. Based on Hauer (1999).
1 Curve radius in feet
Accidents per million vehicle kilometres in horizontal curves with 2 Curve length in feet
radius between 100 and 700 metres
3 Ln(AADT)
Tangent 100 200 300 400 500 600 700 4 Posted speed limit
length 5 Average IRI (International Roughness Index)
(metres) 6 Difference between posted and advisory speed
7 Upstream tangent of 0–600 feet (dummy)
25 0.373 0.298 0.246 0.211 0.187 0.153 0.126
57 0.382 0.304 0.250 0.213 0.187 0.153 0.126 8 Upstream tangent of 601–1200 feet (dummy)
125 0.392 0.309 0.253 0.214 0.187 0.153 0.126 9 Upstream tangent of 1201–2600 feet (dummy)
175 0.402 0.315 0.256 0.216 0.188 0.153 0.126
300 0.428 0.331 0.265 0.219 0.188 0.153 0.126 When applying the equation, curve radius was varied between 328
500 0.473 0.357 0.279 0.226 0.189 0.153 0.126
feet (100 m) and 3937 feet (1200 m). Mean curve radius in the data was
800 0.549 0.400 0.301 0.235 0.190 0.153 0.126
1200 0.671 0.465 0.334 0.249 0.192 0.153 0.126 2920.4 feet and mean curve length was 914.8 feet. Based on this, curve
length was set to a proportion of curve radius = 914.8/2920.4 = 0.313.

2
R. Elvik Accident Analysis and Prevention 133 (2019) 105322

All other variables were entered at their mean values. The mean values Table 2
of the three levels for the length of the upstream tangent, using the Accident rates in curves and on straight sections in Norway. Derived from Eick
midpoint of the range as an estimate and converted to metres was, and Vikane (1992); Eriksen (1993) and Stigre (1993).
respectively, 91 m, 274 m and 579 m. Risk in curves with a radius of Injury accidents per million vehicle
100 m was found to increase sharply as the length of the upstream kilometres
tangent increased.
Mean distance Accident rate in Accident rate on Ratio of accident
The model (Eq. 5) predicts the number of accidents. However, as
between curves curves straight sections rates (Curve/
AADT and the ratio of curve length to curve radius were kept constant (km) straight)
when applying the model, the results can be interpreted as estimates of
accident rate at the mean traffic volume (an AADT of 1338). In the 6.54 0.420 0.081 5.22
3.49 0.675 0.106 6.36
comparisons of accident rates for different distances to upstream
2.52 0.479 0.123 3.89
curves, everything else was kept constant. A curve with radius 1,200 m 1.46 0.188 0.038 4.89
was used as reference for the accident modification factors. 0.89 0.132 0.043 3.05
All 0.410 0.094 4.34

3.4. Bil et al. (2018)

surprising, as higher density of curves means that the straight sections
A paper by Bil, Andrasik, Sedonik and Cicha (2018) presented a GIS-
become shorter, and short straight sections may be associated with a
tool used to identify curves and compute curve radius on Czech high-
lower speed than long straight sections. Nevertheless, the relative in-
ways. The paper contained an accident prediction model for one class of
crease in accident rate (rate in curves/rate on straight section) tends to
road. The authors were contacted and asked if they could supply similar
be smaller on roads with many curves than on roads with few curves,
models developed for other classes of road. The answer was positive,
consistent with what the studies above have found. The following
and prediction models for four classes of road were provided. All these
function was fitted to the relative accident rates:
models had the following form:
L Relative accident rate (curve/straight)=3.652X0.2555 (R2=0.4775)
Number of accidents per curve per year = e β1+ β2 ( R ) AADT β3 L β4 R β5 (6) (7)
The first coefficient (β1) is a constant term. The next coefficient (β2) This function will be applied along with the results of the other
refers to the ratio of the length of a curve to its radius, with both length studies reviewed above.
and radius measured in metres. The final three coefficients refer to
AADT (Annual Average Daily Traffic), Curve length (L) and curve ra- 4. Estimation of risk – sharpness of neighbouring curve
dius (R). Note that the model predicts the number of accidents. To
obtain the accident rate, used as estimator of safety by the other studies Two studies by Gooch et al. (2016, 2018) included data on the
included, the number of vehicle kilometres performed in curves with presence and radius of proximal and distal curves to a given curve.
different radii was estimated using an AADT of 2000 and assuming that Models were developed including these variables in addition to the
each curve had the same length as its radius (L/R = 1). radius of a subject curve. These models were applied to estimate how
One of the four classes of road included in the study was motorways, the accident rate in a subject curve depends on the presence and radius
where the mean distance between horizontal curves was considerable of a proximal curve. The following estimates were developed, all for a
larger than in the other three classes of road. This class was omitted. In subject curve with radius 100 m:
the other three classes of road, the mean number of curves per kilo-
metre of road was 2.4382, 4.9232 and 5.2230, corresponding to mean 1 No proximal or distal curves (intended to represent an isolated
distances between curves of 410, 203 and 191 m. The accident rate in curve).
curves with a radius of 100 m was found to increase as the distance 2 A proximal curve within 0.75 miles with radius 1,200 m.
between curves increased. 3 A proximal curve within 0.75 miles with radius 100 m.
4 A proximal curve within 1.25 miles with radius 1,200 m.
3.5. Re-analysis of Eick and Vikane (1992); Eriksen (1993) and Stigre 5 A proximal curve within 1.25 miles and radius 100 m.
(1993) 6 A proximal curve more than 0.75 miles away with radius 1,200 m.
7 A proximal curve more than 0.75 miles away with radius 100 m.
A number of Norwegian studies (Eick and Vikane, 1992; Eriksen,
1993; Stigre, 1993) evaluated the road safety effects of signing of ha- The distances of less or more than 0.75 miles were used by Gooch
zardous curves. In these curves, background and/or directional signs et al. (2016). The distance of 1.25 miles was used by Gooch et al.
were put up. All curves were identified by means of a computer pro- (2018). The extreme values for radius (1200 and 100 m) were used to
gramme (Amundsen and Lie, 1984). The purpose of this programme probe whether the distance or the sharpness of a proximal curve had the
was to identify surprising curves. Curves scoring high for degree of greatest influence on the accident rate of the subject curve.
surprise were selected for special signing. These curves did not all have
the same radius, but data on the radius of each curve was not provided 5. Results
in the studies quoted above. However, data provided by Sakshaug
(1998) show that the mean radius of the signed curves was 110 m. Fig. 1 shows results of the five studies that were reviewed in section
Thus, no great inaccuracy is introduced by treating the curves as having 3, dealing with how accident rate in a curve with a given radius de-
a radius of 100 m to be consistent with the studies quoted above. pends on the distance to a neighbouring curve. A sharp curve with a
Based on the data provided by Eick and Vikane (1993), Eriksen radius of 100 m is used as case. It is seen that all studies find that ac-
(1993) and Stigre (1993), five groups have been formed with respect to cident rate in a sharp curve increases as the mean distance between
the mean distance between curves. Accident rates (injury accidents per curves increases.
million vehicle kilometre) in these groups are shown in Table 2. The shape of the functions showing how accident rate in a sharp
The estimates of risk in Table 2 show that not only the risk in curves curve increases as the distance to neighbouring curves increases differs
declines as the distance between them gets shorter, but that the risk on considerably. The functions developed by Hauer (1999) and Khan et al.
straight section between curves also declines. This is perhaps not so (2013) rise steeply at an increasing rate. The function developed by

3
R. Elvik Accident Analysis and Prevention 133 (2019) 105322

Fig. 1. Relative risk in horizontal curves with radius 100 m as a function of the mean length of straight section between a pair of curves.

Findley et al. (2012) is close to linear, whereas the functions fitted to 6. Discussion
the Norwegian and Czech studies (Eick and Vikane, 1992; Eriksen,
1993; Stigre, 1993; Bil et al., 2018) increase steeply at first and then Many risk factors that are associated with accidents display a dose-
become flatter. It would therefore not be informative to try to devel- response pattern. The higher the speed, the higher the risk of accident
oped a synthesised function based on the five functions shown in Fig. 1. and the more severe its outcome. The higher the blood alcohol con-
The intercept of the functions also differs considerably. The function centration, the higher the risk of accident. The larger the mass of a
fitted to the study by Bil et al. (2018) suggests a negative accident rate vehicle, the higher its potential for causing damage to others in case of
when the distance between curves goes toward zero; this is implausible, an accident. Horizontal curves appear to display the opposite pattern:
but possibly attributable to the fact that Bil et al. included only curves the more there are of them, the lower the risk in each curve.
in their models, not straight sections. When applying their models, a Based on the studies presented in this paper, one cannot rule out
radius of 1,200 m was treated as a straight section. Had a value of, say that a road with many curves will have a lower total accident rate than
12,000 been applied for a presumably straight section, relative accident a road with few curves. A simple numerical example has been devel-
rate for a radius of 100 m would have been considerably higher. oped to illustrate this. A road section of 1 km with a constant traffic
Despite the rather wide dispersion of intercepts and different func- volume is considered. The section is assumed to have either 7, 5, 3, or 1
tional forms seen in Fig. 1, all studies agree that the more curves there horizontal curves. Each curve has a radius and a length of 100 m. It is
are on a road, the lower is the risk in a curve with a given radius. In assumed that the beginning and end of the road section consists of a
other words: the more common this risk factor is, the lower is the risk curve, except for the case of 1 curve, which is located at the beginning
associated with it. of the road section. Table 3 shows hypothetical relative accident rates
Fig. 2 shows estimates developed on the basis of the two studies by for the curves and the straight sections between them.
Gooch et al. (2016, 2018). The results of the studies were very similar, Accident rate for the shortest straight sections has been given the
but the coefficients for radius (degree of curvature) of proximal curves value of 1. All other accident rates are relative to this value. The section
did not apply the same values for distance in the two studies. The first with seven curves will have six straight sections located between the
study applied distances of less than or more than 0.75 miles from a curves, each with a length of 50 m (0.05 km). For the shortest straight
subject curve. The second study only applied a distance of less than sections, speed is not expected to increase compared to the speed kept
1.25 miles from a subject curve. in curves. For the longer straight sections, an increase in speed has been
A gentle proximal curve within 0.75 miles of a subject curve hardly assumed, leading to an increase in accident rate. Since traffic volume is
influences accident rate in the subject curve. However, if the proximal assumed to be constant throughout the length of the road section, the
curve has a radius of 100 m, accident rate in the subject curve is 8% expected number of accidents can be estimated simply by multiplying
lower (relative risk 0.92). Proximal curves located more than 0.75 miles the length of curves and straight sections by their respective relative
from a subject curve appear to influence accident rate in the subject accident rates. The estimated total number of accidents is shown in the
more than proximal curves located less than 0.75 miles from the subject rightmost column of Table 3.
curve. This is inconsistent with the functions presented in Fig. 1, all of It is seen that the estimated expected number of accidents tends to
which show a positive relationship between distance between curves increase as the number of curves goes down. This obviously follows
and relative risk in a subject curve with a given radius. However, a from the assumptions made, but as these are not altogether implausible,
sharp proximal curve is associated with a reduction in the accident rate the hypothetical estimates may nevertheless predict real data. The re-
of a subject curve for all distances specified by the two studies. It is sults for the two bottom rows of Table 3 show a case of Simpson’s
noted that some of the coefficients estimated by Gooch et al. (2016, paradox. This denotes a situation where an effect in each of two groups,
2018) were highly uncertain and that results could have been different A and B, goes in one direction, whereas the effect when the groups are
with different values for these coefficients added (A + B) goes in the opposite direction. While the accident rate is
higher both in curves and on straight sections in the bottom row than in
the row immediately above it (3.6 vs. 2.7 and 1.3 vs. 1.1), the expected

4
R. Elvik Accident Analysis and Prevention 133 (2019) 105322

Fig. 2. Accident rate in a subject with radius 100 m depending on presence and radius of a proximal curve.

number of accidents is slightly lower than in the next-to-bottom row 7. Conclusions

(1.53 vs. 1.58).
One potential source of bias in comparisons using accident rate, is The main conclusions of the research presented in this paper are:
that accident rate depends on traffic volume and traffic volume could
be different on roads with different frequency of curve. Roads with 1 The shorter the mean distance between horizontal curves, the lower
many curves tend to have low traffic volume and the accident rate tends the accident rate in curves of a given radius.
to be higher at a low traffic volume than at a high traffic volume. 2 Neighbouring curves with a small radius (sharp curves) are asso-
Nevertheless, it is unlikely that differences in traffic volume between ciated with a lower accident rate in a subject curve of a given radius
roads with many curves and roads with few curves can explain the than neighbouring curves with a larger radius.
findings of this paper. Applying the coefficient for ln(AADT) in a recent 3 It cannot be ruled out that, under plausible assumptions, a road with
accident prediction model for Norway (Høye, 2016) (0.928), it can be many sharp curves will have a lower accident rate than an otherwise
estimated that accident rate on a road with an AADT of 1000 will be identical road with fewer sharp curves.
about 18% higher than on an otherwise identical road with an AADT of
10,000. The differences in accident rates associated with horizontal
curve radius and distance to neighbouring curves found in the studies Declaration of Competing Interest
reviewed in this paper are far greater than 18%.
The most likely explanation for the results are behavioural adap- The authors declare that they have no known competing financial
tation among drivers. When driving on a road that mostly consists of interests or personal relationships that could have appeared to influ-
curves, drivers come to expect that there will be many curves. They ence the work reported in this paper.
adapt their speed and visual search accordingly, but not enough to
eliminate the increase in accident rate associated with curves. Even on
roads that mostly consist of curve, the accident rate in the curves re- Acknowledgements
mains higher than on straight sections.
This study was made as part on the ongoing revision of the
Handbook of Road Safety Measures, funded by the Norwegian Ministry
of Transport and the Norwegian Public Roads Administration.

Table 3
Hypothetical accident rates for road sections with different number of curves.
Curves (N) Straight Length in Straight Mean length of Relative accident Relative Expected Expected Total expected
sections (N) curves (km) length (km) straight section rate in curves accident rate accidents in accidents on number of
(km) straight curves straight accidents

7 6 0.7 0.3 0.050 1.5 1.0 1.05 0.30 1.35

5 4 0.5 0.5 0.125 2.0 1.0 1.00 0.50 1.50
3 2 0.3 0.7 0.350 2.7 1.1 0.81 0.77 1.58
1 1 0.1 0.9 0.900 3.6 1.3 0.36 1.17 1.53

5
R. Elvik Accident Analysis and Prevention 133 (2019) 105322

References Hauer, E., 1999. Safety and the choice of degree of curve. Transp. Res. Rec. 1665, 22–27.
Høye, A., 2016. Utvikling av ulykkesmodeller for ulykker på riks- og fylkesvegnettet I
Norge (2010-2015). Rapport 1522. Transportøkonomisk institutt, Oslo.
Amundsen, F.H., Lie, T., 1984. Utforkjøringer kan begrenses. Temahefte 15 i temaserien Khan, G., Bill, A.R., Chitturi, M.V., Noyce, D.A., 2013. Safety evaluation of horizontal
Trafikk. Transportøkonomisk institutt, Oslo. curves on rural undivided roads. Transp. Res. Rec. 2386, 147–157.
Bauer, K.M., Harwood, D.W., 2013. Safety effects of horizontal curve and grade combi- Matthews, L.R., Barnes, J.W., 1988. Relation between road environment and curve ac-
nations on rural two-lane highways. Transp. Res. Rec. 2398, 37–49. cidents. Proceedings of 14th ARRB Conference, Part 4, 105-120.
Bil, M., Andrasik, R., Sedonik, J., Cicha, V., 2018. ROCA – an ArcGIS toolbox for road Montella, A., Galante, F., Mauriello, F., Pariota, L., 2015. Low-cost measures for reducing
alignment and horizontal curve radii computation. PLoS One 13 (12), e0208407. speeds at curves on two-lane rural highways. Transp. Res. Rec. 2472, 142–154.
Christensen, P., Ragnøy, A., 2006. Vegdekkets tilstand og trafikksikkerhet. Rapport 840. Musey, K., Park, S., 2016. Pavement skid number and horizontal curve safety. Procedia
Transportøkonomisk institutt., Oslo. Eng. 145, 828–835.
Eick, H., Vikane, G., 1992. Verknaden av URF-tiltak i Hordaland. Rapport. Statens veg- Persaud, B., Retting, R., Lyon, C., 2000. Guidelines for identification of hazardous
vesen Hordaland, Trafikkseksjonen, Bergen. highway curves. Transp. Res. Rec. 1717, 14–18.
Elvik, R., 2013. International transferability of accident modification functions for hor- Sakshaug, K., 1998. Effekt av overhøyde i kurver: Beskrivelse av datamaterialet. Notat av
izontal curves. Accid. Anal. Prev. 59, 487–496. 2.11.1998. SINTEF, Bygg og miljøteknikk., Trondheim.
Elvik, R., 2017. Can evolutionary theory explain the slow development of knowledge Saleem, T., Persaud, B., 2017. Another look at the safety effects of horizontal curvature on
about the level of safety built into roads? Accid. Anal. Prev. 106, 166–172. rural two-lane highways. Accid. Anal. Prev. 106, 149–159.
Eriksen, T., 1993. Analyse av utforkjøringsulykker i Akershus fylke 1987-92. Stigre, S.A., 1993. Tiltak mot utforkjøringsulykker i vestfold. Effektundersøkelse.
Hovedoppgave i samferdselsteknikk. Norges Tekniske Høgskole, Institutt for Oppdragsrapport til Statens vegvesen Vestfold. Svein A. Stigre., Rykkinn.
Samferdselsteknikk, Trondheim. Tom, G.K.J., 1995. Accidents on spiral transition curves. ITE-Journal, September 1995,
Findley, D., Hummer, J.E., Rasdorf, W., Zegeer, C.V., Fowler, T.J., 2012. Modeling the 49–53.
impact of spatial relationships on horizontal curve safety. Accid. Anal. Prev. 45, Zegeer, C., Stewart, R., Reinfurt, D., Council, F., Neuman, T., Hamilton, E., Miller, T.,
296–304. Hunter, W., 1991. Cost-Effective Geometric Improvements for Safety Upgrading of
Gooch, J.P., Gayah, V.V., Donnell, E.T., 2016. Quantifying the safety effects of horizontal Horizontal Curves. Report FHWA-RD-90-021. US Department of Transportation,
curves on two-way two-lane rural roads. Accid. Anal. Prev. 92, 71–81. Federal Highway Administration, Turner-Fairbank Highway Research Center,
Gooch, J.P., Gayah, V.V., Donnell, E.T., 2018. Safety performance functions for horizontal McLean, VA.
curves and tangents on two lane, two way rural roads. Accid. Anal. Prev. 120, 28–37.