Safety effects of section control - An empirical Bayes evaluation

Accident Analysis and Prevention 74 (2015) 169–178

Contents lists available at ScienceDirect

Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap

Safety effects of section control - An empirical Bayes evaluation Alena Høye * Institute of Transport Economics, Department of Safety, Security and Environment, Gaustadalleen 21, 0349 Oslo, Norway

A R T I C L E I N F O

A B S T R A C T

Article history: Received 24 July 2014 Received in revised form 13 October 2014 Accepted 14 October 2014 Available online xxx

The safety effects of section control were investigated at 14 sites in Norway. A before–after study was conducted with the empirical Bayes method in order to control for regression to the mean (RTM). Effects of trend, volumes, speed limit changes and speed cameras at some of the sites in the before period are controlled for as well. For injury crashes a non-significant reduction by 12% was found. The number of killed or severely injured was found to be significantly reduced by 49% at the section control sites. The results indicate that the crash reductions in tunnels (most of which are undersea tunnels with section control on steep downhill segments) are at least of the same magnitude as on open roads. The results are consistent with findings from speed measurements, although the crash reductions are larger than one would expect as a result from the speed reductions. Downstream of the section control sites (up to 3 km in each direction) injury crashes were found to be significantly reduced by 46%. The number of KSI downstream of the section control sites is too small for drawing any conclusions. It is concluded that section control is effective in reducing both speed and crashes, especially serious crashes, and that spillover effects (crash reductions at non-enforcement sites) are more likely to occur than crash migration. The size of the effects that were found should be interpreted with caution because of the relatively short after periods for some of the sites and the sensitivity of the results to the outcomes of individual crashes. ã 2014 Elsevier Ltd. All rights reserved.

Keywords: Empirical Bayes Section ?control Enforcement Regression to the mean Crash migration Spillover effects

1. Introduction Section control aims at reducing speed and thereby crashes, especially the most serious crashes. A section control site consists of a stretch of road between two speed cameras (four speed cameras at sites with bidirectional section control), both of which take pictures of all passing vehicles in one direction with automatic license plate recognition. In Norway, data can be saved only from vehicles with an average speed above the speed limit. Average speed is calculated from the times at which the first and second camera is passed. Speed at the camera sites is not taken into account. Drivers of vehicles with an average speed above the speed limit may be prosecuted. In contrast to conventional speed cameras where it is easy to see if the camera takes a picture or not, drivers cannot see whether offences will be prosecuted or from what threshold. All section control sites are signposted at some distance upstream to the first camera, and all cameras are well visible. While speed cameras have been evaluated in a larger number of studies, only four studies were fund that have empirically investigated the effects of section control on crashes

* Tel.: +47 48898074. http://dx.doi.org/10.1016/j.aap.2014.10.016 0001-4575/ ã 2014 Elsevier Ltd. All rights reserved.

(Brassøe et al., 2011; Broughton et al., 2012; Montella et al., 2012Stefan and Winkelbauer, 2005). These studies found reductions of total crash numbers between 24% and 52% (on average about
30%) and a reduction of the number of severe crashes by 56%. However, most of these studies have not controlled for regression to the mean (RTM) and it is therefore possible that at least some of the results are overestimated. In Norway, section control has been installed at two sites in 2009 and at 18 additional sites by 2013. About half of all section control sites are in tunnels, most of which are undersea tunnels. Most section control sites in tunnels are between the entrance and the bottom of the tunnel in the downhill direction. Section control is regularly only installed on roads with exceptionally unfavorable crash records in Norway, except in tunnels where section control mostly was installed because of a lack of suitable sites for regular police enforcement with radar or laser devices. Section control is quite controversial in Norway, as in other countries (Tay, 2010; Wilson et al., 2012). Opponents are concerned about privacy issues and have doubts about its effectiveness, especially in tunnels where high crash records are not required for installing section control. Additionally, crash migration effects are expected. Crash migration occurs when crashes are avoided at one site, but occur at other sites instead.

5 11 19 30.39

6

5 11 19 30.39

6

0 0 0 0 5 0 3 4 2 0 1 1 1 1 1 0 2 1 3 6 4 1 3 2 6.00 6.00 6.00 3.39 6.00 3.00

2 0 3 3 1 1 0 4 1 1 0 0 1 0 17 5 12 2 5 2 7 3 2 1 6 2 1 0 5 1 1 38 17 21 8.08 7.98 21.44 20.72 15.18 11.66 27.61 81.07 10.46 2.95 4.02 32.26 12.22 12.22 267.88 90.20 177.68

0 2 2 3 2 1 1 1 1 5 0 2 1 1 22 12 10

Before After Before

Directions with section control (one or both)

Oslofjord W Oslofjord E Barkald Finstad Rosten Harestua Bromma Dørdal Byfjord N Eiksund S Eiksund N/S Hell N Troms S Troms N All Tunnels Open roads

One tube One tube No No No No No No One tube One tube One tube One tube Two tubes Two tubes

One One Both Both Both Both Both One One One Both One One One

2.49 2.36 9.59 10.54 4.18 3.72 7.85 8.30 2.65 2.98 4.07 4.19 2.05 2.05 67.02 22.84 44.18

16 16 36 15 36 13 29 36 16 18 18 18 20 20 21.9 17.8 27.5

6,669 6,952 2,042 4,310 3,317 7,924 3,987 8,920 8,113 1,808 1,803 14,062 9,803 9,799 5,670 7,283 4,837

6,953 7,196 1,904 4,452 3,483 8,126 4,103 9,205 8,533 2,155 2,153 15,089 10,253 9,818 5,894 7,726 4,948

18.18 17.97 21.44 49.74 15.18 32.28 34.27 81.07 23.54 5.90 8.04 64.52 22.00 22.00 416.12 182.14 233.98

2. Data

a

Before

Speed reducing measures may cause crash migration when drivers speed up after having passed through a control section or as a consequence of traffic diversion (Decina et al., 2007; Mountain et al., 2004). In Norway, crash migration because of route diversion is unlikely, most section control sites have no realistic alternative routes. Crash migration because of increasing speed on dowstream sections is however possible. Supporters on the other hand claim that section control may have positive effects especially on steep downhill sections in tunnels where speed often is high, and that section control may have spillover, rather than crash migration effects on downstream sections. Spillover effects occur when crash reductions occur not only at those sites where a measure is implemented, but also at other sites. Section control may cause spillover effects when drivers keep a low speed after having passed a control section (instead of accelerating). The aim of the present study is to conduct a thorough investigation of the effects of section control on casualty crashes and on the number of killed or severely injured (KSI), including possible crash migration/spillover effects on downstream sections, to compare the effects of section control between tunnels and open roads, and to compare the crash effects of section control to the speed effects that were found in previous studies. Since section control often is installed at highcrash locations, results from before-after studies may be affected by regression to the mean (RTM) unless a comparison group of similar sites with similar crash records in the before period as the section control sites is used, or RTM is controlled for by other means. A comparison group of similar roads with similar crash records in the before period was not available. Therefore an empirical Bayes (EB) evaluation was conducted that controls for RTM by comparing observed to expected crash numbers in the after period.

0 0 0 0 0 2 0 0 1 0 0 0 0 0 3 1 2

After Before

KSI Injury crashes

Before Before After

Length(km) Dir.a Tunnel

Table 1 Description of section control and downstream sites.

Months

AADT

After

Mill. vehicle km

After

After

Down-streamlength (km)

Injury crashes downstream

After

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

KSI downstream

170

All sites with section control in Norway were included in the evaluation where section control was installed in 2012 or earlier. Thus, 14 sites were available for the evaluation. Crash data was available until 2013. For those sites that were installed in 2013 (six sites) crash data was available for less than one year and these sites were therefore not included in the evaluation. For those sites that are included in the evaluation some basic road characteristics, as well as crash numbers in the before and after period, are summarized in Table 1. The speed limit is 80 km/h at most sites (except Barkald and Finstad where it was 90 km/h in the first part of the before period, and Oslofjord tunnel where it was reduced from 80 to 70 km/h at the end of the before period). The number of lanes is two at most sites (three at three of the tunnel sites). All sites are outside urban areas. All tunnels are undersea tunnels, except Hell N. The before period was 36 months for all sites, the after period was between 13 and 36 months (21.9 months on average). All after periods started three months after the cameras were installed and ended in December 2013 or after 36 months. All before periods are three whole years, ending in December before the year of installation. Traffic volumes have increased from the before to the after period at all except one section control sites. The average increase (all sites) was by 4.4%. The five rightmost columns of Table 1 describe the downstream sections that were used in the investigation of possible crash migration or spillover effects. These were only investigated on open roads. The downstream sections consist of a maximum length of 3 km downstream of the section control sites. Therefore, bidirectional sites have for the most part 6 km of downstream length and the one site with section control in only one direction has 3 km of downstream length. The site with 3.39 km of

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

downstream length has a tunnel about 400 m after the second speed camera in one direction. All downstream sites have the same general road characteristics as the section control sites (same speed limit, number of lanes, AADT, etc.). 3. Method The effect of section control on injury crashes and on the number of KSI was investigated in a before–after evaluation, both at the section control sites and at the downstream sites. The main part of the evaluation has been conducted with the EB method which is regarded as suitable for controlling for RTM in before– after studies (Elvik, 2008). Effects of section control were also calculated with a similar design but without control for RTM, and as a simple comparison of crashes per million vehicle kilometers travelled in the before and after period. Additionally, the effects of section control on injury crashes and KSI that were found in the before-after EB study are compared with the effects of section control on speed that were found in previous studies. 3.1. EB before–after evaluation According to the EB procedure, the observed number of crashes on roads with section control in the after period (Oa) is compared to the expected number of crashes on the same roads in the after period (Ea), i.e., the number of crashes that would have occurred without section control. An unbiased estimate of the effect of section control is according to Hauer (1997): Effect ¼

Oa =Ea 1 þ VarðEa Þ=E2a

171

The estimated percentage change of the number of crashes from the before- to the after period is (Effect-1)  100. Ea is for each site estimated as a function of the expected number of crashes in the before period (Eb) and changes over time. Eb is a function of the crash history, i.e., the observed number of crashes in the before period (Ob), the predicted number of crashes in the before period (Pb), and a statistical weight (w): Eb ¼ w  Pb þ ð1
wÞ  Ob Pb is the number of crashes that would have been expected in the before period on road sections with the same AADT and other characteristics as the one with section control. The safety performance function (SPF) that is used to estimate Pb is described by Høye (2014a). It is based on crash prediction models for the major part of the Norwegian road network (except private and municipal roads and tunnels). The models are negative binomial models with a variable overdispersion parameter. Since about half of all section control sites are in tunnels, new models were developed with the same database and predictor variables as the original models, but tunnels were added to the database and several predictor variables were defined for tunnels (Høye, 2014b). The coefficients for all predictor variables that are relevant for at least one of the section control sites are shown in Table 2 for injury crashes and for the number of KSI, both in the original model and in the tunnel model. Section length and the number of data years were included as exposure variables, i.e., the coefficients for the natural logarithm of these variables were set to one in the crash prediction models. Coefficients for the overdispersion parameter are shown in the last three rows of Table 2. Predictor variables that are not relevant are not shown (e.g., none of the sites has a speed limit of 100 or below 80 km/h and coefficients for other speed limits than 80 and 90 km/h are therefore omitted from Table 2).

Table 2 Generalized negative binomial model parameter estimates (standard deviations) (reference). Original models

Tunnel models

Injury crashes

KSI

Injury crashes

KSI

Ln(AADT) Ln(AADT)2 Speed limit: 80 km/h (dummya )

1.230 (0.066)
0.016 (0.004) (reference)

1.937 (0.167)
0.074 (0.011) (reference)

1.214 (0.066)
0.015 (0.004) (reference)

1.933 (0.165)
0.074 (0.011) (reference)

Speed limit: 90 km/h (dummya ) Speed limit: 90 or 100 km/h (dummya ) N of lanes: 2 (dummya ) N of lanes: 3 (dummya ) T-junctions (numberb) X-junctions (numberb) Curves (80 km/h) (numberb) Curves (90 or 100 km/h) (numberb) Vertical grades (80 km/h) (numberb) Vertical grades (90 or 100 km/h) (numberb) Vertical grades (numberb) Road category: 2-/3 lane with median barrier (dummya ) Road category: Trans European network road (dummya ) Road category: other national road (dummya ) Road category: district road (dummya ) County (one dummy for each of 19 counties) Tunnel (dummya ) Two-tube tunnel (dummya ) Undersea tunnel (dummya ) Short tunnel (< 1 km) (dummya ) Constant Ln (km  years) Overdispersion parameter Ln (km  years) AADT Constant


0.248 (0.070)

a

(reference) 0.110 (0.062) 0.138 (0.016) 0.314 (0.041) 0.218 (0.019) 0.183 (0.251)
0.005 (0.033) 0.496 (0.232)


0.291 (0.069)
0.343 (0.159) (reference) 0.160 (0.157) 0.111 (0.043) 0.176 (0.121) 0.120 (0.044) 0.034 (0.568)

(reference) 0.094 (0.061) 0.142 (0.016) 0.312 (0.041) 0.216 (0.019) 0.226 (0.250)
0.006 (0.033) 0.439 (0.228)


0.383 (0.159) (reference) 0.191 (0.153) 0.113 (0.043) 0.167 (0.121) 0.120 (0.044) 0.074 (0.568)


0.117 (0.059)
0.123 (0.163) 0.259 (0.063) 0.220 (0.057) (reference)


0.440 (0.065)
0.106 (0.027)
0.029 (0.023) (reference)


0.124 (0.059)
0.120 (0.153) 0.246 (0.062) 0.197 (0.057) (reference)


17.62 (0.254) 1.000


21.05 (0.628) 1.000


0.303 (0.091)
0.064 (0.119)
0.407 (0.232) 0.326 (0.114)
17.56 (0.251) 1.000

0.001 (0.185)
0.353 (0.331)
0.217 (0.45)
0.027 (0.289)
21.03 (0.622) 1.000


0.778 (0.028)
0.371 (0.035) 8.53 (0.439)


0.787 (0.062)
0.730 (0.047) 13.70 (0.734)


0.775 (0.028)
0.368 (0.034) 8.49 (0.43)


0.793 (0.061)
0.728 (0.046) 13.75 (0.723)


0.499 (0.070)
0.101 (0.027)
0.024 (0.023) (reference)

Dummy variables: 1 if yes, 0 otherwise. Number of junctions, curves etc.: Ln(number of X + 1), for curves separate variables are defined for each speed limit, for vertical grades separate variables are defined for each speed limit in the models for injury crashes. b

172

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

3.2. Before–after studies without control for RTM In order to estimate the effects of section control without controlling for RTM, crash counts were compared between the before and after period, while taking into account differences in period length, traffic volumes and general changes of crash risk between the before- and after period. The effect of section control on crashes is estimated as Effect ¼

Oa =Ob Pa =Pb

where Oa,Ob, Pa, and Pb are defined as described above. Thereby, the same factors are controlled for as in the EB evaluation, except RTM. The estimated effect of section control without control for RTM or other factors (except period length and traffic volumes) is Effect ¼

Oa Mill:vechilekma = Ob Mill:vechilekmb

2012

2011

0.0

2010

Relave N of KSI

2009

0.2

2008

Relave N of injury crashes

2007

0.4

0.68 0.67

2013

0.82

0.75 0.73

0.6

0.80

0.94

0.88 0.86

0.8

0.93

1.00 1.00

1.0

1.06

1.2

1.07

1.4 1.15

The overdispersion parameter and the statistical weights become smaller with increasing predicted crash numbers, and the expected crash numbers are therefore closer to the observed crash numbers, the higher the predicted crash numbers. Eight of the 14 sites with section control had speed cameras in the whole or a part of the before period. Since speed cameras are not a predictor in the crash prediction models, the expected crash numbers for those sites with speed cameras in the before period were adjusted for crash effects of speed cameras. It was assumed that speed cameras affect crashes on a length of road of one kilometer, and the average effect of speed cameras was assumed to be a reduction of injury crashes by 20% and KSI crashes by 30%. These effects are based on a meta-analysis of the safety effects of speed cameras by Høye (2014c). In the metaanalysis, speed cameras were found to reduce injury crashes by 20% on average and fatal crashes by 51%. The effect on fatal crashes has been adjusted downwards to 30% because the result for fatal crashes may be affected by RTM and because the effect on the number of KSI is assumed to be smaller than the effect on fatal crashes. Speed limit changes at four of the sites in the before period were taken into account by adjusting the expected crash numbers in the before period for the assumed effect of the speed limit changes on crashes. It was assumed that speed was reduced by 3.6 km/h which is the average effect of speed limit changes by 10 km/h according to a meta-analysis by Elvik (2012). The effects on injury crashes and KSI of a speed reduction by 3.6 km/h were estimated with the help of the power model by Elvik (2009). Accordingly, the speed limit reductions from 90 to 80 km/h and from 80 to 70 km/h were estimated to have reduced injury crashes by 6.3% and 7.1% respectively, and the number of KSI by 14.1% and 15.7%, respectively. Changes over time are taken into account by applying a trend factor that has been developed along with the crash models by Høye (2014a). Without adjustment, the model predictions refer to the year 2008. Crash risk has however decreased considerably over time. A trend factor has therefore been developed that takes into account general changes of traffic volumes and crash numbers in Norway over time. It allows the conversion of model predictions to any year between 1997 and 2020. The relative numbers of injury crashes and KSI in the years 2006–2013 according to the trend factor are shown in Fig. 1. For example from 2008 to 2012 the number of KSI per million vehicle kilometers has decreased by 27% (1–0.73) and the predicted number of KSI in 2012 is therefore calculated as 0.73 times the predicted number of KSI in 2008. According to the trend factor, the general decrease of crash and injury risk over time would have contributed to a decrease of the annual numbers of injury crashes by 19.4% from the before- to the after period and to a decrease of the number of KSI by 19.9% from the before- to the after period if all else had remained equal.

1.11

1 1 þ Pb =’

2006

w¼

The different lengths of the before- and after periods, changes of traffic volumes from the before- to the after period and the (nonlinear) relationship between traffic volumes and crash numbers, as described by the SPF (see above), are taken into account as well. In order to calculate aggregated effects for a number of sites, the numbers of Oa and Ea are summed up over all sites and the summary effect is calculated with the formula described above. Standard deviations and confidence intervals for aggregated effects were calculated as described by Hauer (1997) and Persaud et al. (2005). The comparison of the observed crash numbers in the after period with expected crash numbers is the part of the EBprocedure that is “responsible” for the control for RTM. In cases where the observed crash numbers have been exceptionally high in the before period, the expected crash numbers in the before period are adjusted downwards. Thus, the base of comparison is assumed to be closer to the long-term mean crash numbers, and the results will not (or to a lesser extent) be overestimated due to RTM.

Rel. N of injury crashes / KSI (1 i 2008)

Omitted predictor variables are dummy variables for speed limits 30, 40, 50, 60, 70, and 100, four, five and six or more lanes, motorway, median, median barrier, median, and midline rumble strips, as well as numbers of roundabouts, off-ramps, on-ramps, curves (separate variables at different speed limits), and vertical grades (separate variables at different speed limits). Coefficients for the 19 county dummy variables are also omitted. The statistical weight (w) is a function of the overdispersion parameter f that is estimated along with Pb with the SPF and Pb (Hauer et al., 2002):

Fig. 1. Relative numbers of injury crashes and KSI in 2006–2013 according to the trend factor.

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

173

Table 3 Estimated effects of section control on injury crashes (results from EB-evaluation). Observed

Oslofjord W Oslofjord E Barkald Finstad Rosten Harestua Bromma Dørdal Byfjord N Eiksund S Eiksund N/S Hell N Troms S Troms N All Finstad omitted Open roads Finstad omitted Tunnels

Predicted

Expected

Before

After

Before

After

Before

After

2 5 2 7 3 2 1 6 2 1 0 5 1 1 38 31 21 14 17

2 0 3 3 1 1 0 3 1 1 0 0 1 0 16 13 11 8 5

0.72 0.46 2.03 12.79 1.86 1.68 3.27 6.02 0.86 0.31 0.41 2.64 0.89 0.97 34.89 22.10 27.64 14.85 7.25

0.25 0.16 1.45 4.27 1.53 0.48 2.11 4.84 0.32 0.15 0.13 1.13 0.41 0.43 17.65 13.38 14.68 10.41 2.97

1.72 3.58 2.00 7.03 2.91 1.98 1.06 6.00 1.80 0.67 0.14 4.91 0.98 0.99 35.79 28.75 20.99 13.95 14.80

0.59 1.22 1.42 2.35 2.40 0.57 0.69 4.83 0.67 0.32 0.04 2.10 0.46 0.45 18.10 15.75 12.26 9.91 5.85

4. Results 4.1. Before–after evaluation of the crash effects of section control at the section control sites The results of the EB-evaluation are summarized in Table 3 for injury crashes and in Table 4 for KSI. Tables 2 and 3 show for each site, as well as for all sites, tunnels and open roads, the observed crash numbers (Ob and Oa), predicted crash numbers (Pb and Pa) and expected crash numbers (Eb and Ea), as well as the estimated percentage changes of crash numbers with 95% confidence intervals (CI). No CI are shown for individual sites (most CI for individual sites are either very large, ranging over more than one hundred percentage points, or could not be calculated because of zero crashes in the after period). The overdispersion parameters for the individual sites, calculated with the crash prediction models as described above, and the EB-weights are shown as well. The results for injury crashes (Table 3) show that crash counts were 9% above the predictions on average. Six of the sites had more injury crashes in the before period than predicted (when the predicted crash numbers are rounded to the nearest whole number). One of the sites had far fewer injury crashes than predicted (Finstad). Without Finstad, the total injury crash counts were 40% higher than predicted. In order to test the influence of this site on the overall results, these have been calculated without Finstad in addition to the results that are based on all sites. In tunnels the injury crash counts were 138% higher than predicted, on open roads they were 24% lower than predicted (6% lower without Finstad). Overall, injury crashes were found to be reduced by 12% (by 19% when Finstad is omitted) and the results indicate that section control has greater effects in tunnels than on open roads. None of the results is statistically significant. However, a reduction of injury crashes by 12% (or by 19%) is unlikely to become statistically significant unless the number of sites is at least ten times as large as it is in the present evaluation (or at least five times as large without Finstad)1. The evaluation procedure has in other words only weak statistical power.

1 In the absence of a standardized procedure for calculating statistical test power, effects were calculated based on two, three, four etc. times the available section control sites, until the effects became statistically significant. Whether or not effects are statistically significant depends however not only on the total number of sites (or crashes) included in the evaluation, but also on the distribution of observed, predicted, and expected crash counts.

Overdisp.-parameter

EB-weight

Percent change (95% CI)

0.203 0.209 0.101 0.077 0.157 0.137 0.096 0.064 0.180 0.286 0.225 0.103 0.205 0.205

0.221 0.313 0.048 0.006 0.078 0.076 0.028 0.010 0.173 0.476 0.357 0.038 0.188 0.175

+163
100 +72 +25
62 +43
100
19 +19 +10 –
100 +17
100
12 (
34; +9)
19 (
41; +4)
12 (
31; +18)
21 (
5%; +13)
17 (
54; +20)

The results for KSI (Table 4) show that there were about three times as many KSI in the before period than predicted, about six times as many in tunnels and about twice as many outside tunnels. At one of the sites (Eiksund S) there were five KSI in the before period (all five were fatalities in a single crash). When this site is omitted, there were still more than twice as many KSI in the before period than predicted and 3.7 times as many in tunnels. In order to test the influence of this site on the overall results, these have been calculated without Eiksund S in addition to the results that are based on all sites. The overall effect of section control on the number of KSI is a statistically significant reduction by 49%. The result changes only marginally when Eiksund S is omitted. In tunnels and on open roads similar effects were found, and most results are statistically significant, despite the small numbers of KSI that are available for the evaluation. The results for KSI remain unchanged if the number of KSI in Eiksund S in the before period is set to one instead of five (not shown in Table 4). In order to compare the results from the EB evaluation with the results from the evaluation without control for RTM and with control for vehicles miles travelled only, Figs. 1 and 2 summarize the results that were obtained with all three methods for injury crashes and KSI respectively. Table 5 shows the effects of section control on injury crashes and KSI that were estimated without control for RTM and with control for vehicles miles travelled only. The results in Table 5 and Figures 2 and 3 show that the estimated effects of section control are consistently greater when estimated with the EB-procedure than without control for RTM (results “Without control for RTM” are based on a before-after study design that controls for the same factors as the EB-evaluation, except RTM), except for injury crashes on open roads where the observed number of crashes had been smaller than predicted in the before period. The differences between the EB-results and the results without control for RTM are far larger (in terms of percentage points difference) for KSI than for injury crashes, and larger in tunnels than on open roads. In order to investigate the effects of some of the assumptions that were made in the calculation of predicted crash and KSI numbers in the before period, effects of section control were calculated without adjusting the predicted crash/KSI numbers in the before period for the assumed effects of speed limit changes at four of the sites, and without adjustment for the assumed effects of speed cameras at eight of the sites in the before period. Without adjustment for the assumed effects of speed limit changes, the

174

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

Table 4 Estimated effects of section control on numbers of KSI (results from EB-evaluation). Observed

Oslofjord W Oslofjord E Barkald Finstad Rosten Harestua Bromma Dørdal Byfjord N Eiksund S Eiksund N/S Hell N Troms S Troms N All Eiksund S omitted Open roads Tunnels Eiksund S omitted

Predicted

Expected

Before

After

Before

After

Before

After

0 2 2 3 2 1 1 1 1 5 0 2 1 1 22 17 10 12 7

0 0 0 0 0 2 0 0 1 0 0 0 0 0 3 3 2 1 1

0.25 0.26 0.55 1.10 0.78 0.56 1.27 1.06 0.32 0.11 0.13 0.66 0.16 0.18 7.41 7.30 5.32 2.09 1.97

0.08 0.08 0.38 0.35 0.61 0.16 0.81 0.81 0.12 0.05 0.04 0.28 0.08 0.08 3.93 3.88 3.13 0.80 0.75

0.22 0.53 1.04 2.30 1.23 0.73 1.10 1.02 0.46 0.27 0.13 1.39 0.26 0.28 10.95 10.68 7.43 3.53 3.25

0.07 0.16 0.72 0.74 0.96 0.21 0.70 0.78 0.17 0.12 0.04 0.58 0.12 0.13 5.50 5.38 4.12 1.39 1.26

Overdisp.-parameter

EB-weight

Percent change (95% CI)

1.461 1.479 1.055 0.635 1.350 0.924 0.800 0.382 1.206 3.278 2.565 0.562 1.288 1.288

0.853 0.849 0.658 0.367 0.633 0.621 0.387 0.265 0.793 0.967 0.951 0.458 0.887 0.877

–
100
100
100
100 329
100
100 31
100 –
100
100
100
49 (
81;
18)
48 (
80;
16)
55 (
96;
14)
53 (
100;
3)
48 (
100; +8)

and one site had five KSI in one of the two injury crashes, which was a head-on collision. Had there only been two KSI in this crash, the effect of section control would have been an almost statistically significant reduction by 48% (95% CI [
97; +1]). Had there been three KSI in this crash, the effect of section control would have been a non-significant reduction by 22% (95% CI (
95; +51)). Thus, the result is highly sensitive for the outcome of this one crash. In order to compare the results from the EB evaluation with the results from the evaluations without control for RTM Fig. 4 shows the effects of section control on injury crashes and KSI on downstream sections that were obtained with all three methods. The only result that is statistically significant is the one for injury crashes from the EB evaluation. The result without control for RTM (which is controlled for the same factors as the result from the EB evaluation, except RTM) is almost the same as the result from the EB evaluation for injury crashes (
47%, 95% CI (
80; +43)). The observed number of injury crashes in the before period was 30% higher than predicted (Table 6), but the expected number of injury crashes in the before period is almost as high as the observed

results of the EB evaluation differ by between 0.1 and 1.2 percentage points from the original effects (with adjustment for speed limit changes). Without adjustment for the assumed effects of speed cameras, the results of the EB evaluation differ by between 0.0 and 0.9 percentage points from the original effects. Whether or not the results are statistically significant does not change for any of the results without these adjustments. Thus, the adjustments for the assumed effects of speed limit changes and speed cameras have contributed only to a small degree to the results from the EBevaluation. 4.2. Before–after evaluation of the effects of section control on downstream sections The results of the EB-evaluation on downstream sections are summarized in Table 6 for injury crashes and in Table 7 for KSI. For injury crashes, a statistically significant reduction by 46% was found. KSI were found to increase by 30% according to the EB evaluation, this result is however far from being statistically significant. Moreover, most sites had no KSI in the after period,

0 -5 -10 -15

-10 -12

-12 -16

-17

-20

-19 -21

-22

-25

-20

-25

-30 -32

-35

-34 -36

-40

-40 -43

-45 -50 All sites

All sites (Finstad omied) EB

Tunnels

Without control for RTM

Open roads

Open roads (Finstad omied)

Control for veh. km only

Fig. 2. Effects of section control on injury crashes (percentage changes), estimated with different methods. None of the results are statistically significant.

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

175

Table 5 Estimated effects of section control on numbers of KSI, results from before–after evaluation without control for RTM. Injury crashes

All Outliers omitted Open roads Finstad omitted Tunnels Eiksund S omitted

KSI

Without control for RTM

Control for vehicles miles travelled only

Without control for RTM

Control for vehicles miles travelled only


16 (
57; +64)
22 (
62; +62)
10 (
60; +102)
20 (
68; +104)
25 (
76; +139)


36
40
32
34
43


73 (
92;
2)
65 (
91; +29)
66 (
93; +72)


79 (
94;
31)
75 (
93;
13)
74 (
94; +18)


81 (
98; +65)
65 (
96; +246)


84 (
98; +24)
72 (
97; +126)

(
62; (
68; (
67; (
72; (
79;

+16) +18) +40) +57) +54)

0 -10 -20 -30 -40 -50

-48*

-49*

-48 -53*

-55*

-60 -65

-70 -73*

-80

-65 -75

-79*

-81

-90 All sites

All sites (Eiskund S omied) EB

-66 -72

-84

Tunnels

Without control for RTM

-74

Tunnels (Eiksund S omied)

Open roads

Control for veh. km only

Fig. 3. Effects of section control on KSI (percentage changes), estimated with different methods. Statistically significant results are marked with asterisks.

number because of the small statistical weight given to the predicted number. Therefore, the adjustment for RTM was only small. For KSI a considerably more favorable effect was found without control for RTM than in the EB evaluation (
24%, 95% CI (
76; +144)). A large regression effect was expected because the observed number of KSI in the before period was about three times as high as predicted (Table 7). When only the numbers of vehicle miles travelled are controlled for the results are still more favorable:
58% (95% CI (
83; +6)) for injury crashes and
39% (95% CI (
79; +76)) for KSI.

Table 7 shows additionally the effects on injury crashes and KSI that could have been expected as a result of the speed reductions, when taking into account the general relationship between speed and crashes. The latter is described by the power model (Elvik, 2009; Cameron and Elvik, 2010), according to which crash numbers on average change as following as a function of the change of average speed:

4.3. Comparison of crash and speed effects of section control

Crashesa and Crashesb are crash numbers with average speed a and b, Speeda and Speedb are the average speeds a and b, and Exponent has been estimated empirically for different types of road (urban roads, rural roads, motorways, and all roads) and for different degrees of severity. The results in Table 8 were calculated with the exponents for rural roads. According to these results the

For eight of the section control sites, five of which are in tunnels, results from speed measurements before and after the installation of section control are available (Ragnøy, 2011, 2013). The results from the speed measurements are summarized in Table 8.

Crashesa ¼ Crashesb



Speeda Speedb

Exponent

Table 6 Estimated effects of section control on injury crashes on downstream sections (results from EB-evaluation). Observed

Barkald Finstad Rosten Harestua Bromma Dørdal All

Predicted

Expected

Before

After

Before

After

Before

After

3 6 4 1 3 2 19

1 1 1 0 2 1 6

1.27 5.77 2.65 0.75 3.18 0.94 14.56

0.91 1.93 2.19 0.22 2.05 0.76 8.04

2.82 6.00 3.94 0.96 3.01 1.86 18.59

2.01 2.00 3.25 0.28 1.94 1.50 10.97

Overdisp.-parameter

EB-weight

Percent change (95% CI)

0.146 0.119 0.118 0.148 0.118 0.141

0.103 0.020 0.043 0.165 0.036 0.130


55
51
71
100
6
47
46 (
64;
29)

176

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

Table 7 Estimated effects of section control on numbers of KSI on downstream sections (results from EB-evaluation). Observed

Barkald Finstad Rosten Harestua Bromma Dørdal All

Predicted

Expected

Before

After

Before

After

Before

After

3 4 2 0 1 1 11

0 0 0 0 5 0 5

0.34 0.61 0.99 0.30 1.12 0.26 3.63

0.24 0.20 0.77 0.09 0.71 0.20 2.21

0.83 1.91 1.49 0.23 1.06 0.44 5.95

0.57 0.62 1.17 0.07 0.67 0.34 3.43

30

40 20

EB

0 Without control for RTM

-20 -24

-40 -60

-46* -47

-80

-39 -58

Control for vehicles miles travelled only

-100 Injury

KSI

Fig. 4. Effects of section control on KSI (percentage changes) on downstream sections, estimated with different methods. Statistically significant result marked with an asterisk.

speed reductions would have been expected to reduce the number of injury crashes by 5–6% and the number of KSI by 19–20% which is considerably less than the effects that were found in the crash evaluation. 5. Discussion of the results The results from the EB-evaluation of section control indicate that the number of KSI is statistically significantly reduced by 49% and that the number of injury crashes is reduced by 12% on average at all section control sites. The latter effect is not statistically significant, but the test has only weak statistical power and a far larger number of crashes would be required in order to obtain a statistically significant result of this magnitude. These results refer to all section control sites, independent of whether or not they are in tunnels, in what type of tunnels and whether or not section control is uni- or bidirectional. Since the expected crash numbers in the after period are adjusted for the assumed effect of speed cameras at some of the sites in the before period, the results refer to the effect of section control at sites without fixed speed cameras in the before period.

Overdisp.-parameter

EB-weight

Percent change (95% CI)

1.526 0.989 1.016 0.994 0.989 0.852

0.816 0.617 0.507 0.766 0.469 0.763


100
100
100 – 423
100 +30 (
92; +125)

On road sections up to 3 km downstream of the section control sites the number of injury crashes was found to be reduced significantly by 46% and for KSI an increase by 30% was found that is far from being statistically significant. There are several factors that should be taken into account in interpreting the results for the section control sites. These are discussed in the following. Firstly, there were two sites where the observed crash and KSI numbers in the before period were far above/below the predicted crash or KSI numbers. Secondly, the effects of RTM may be overestimated in tunnels. Thirdly, differences between sites with uni- and bidirectional section control should be addressed. And finally, the differences between those crash effects that were found empirically and those that would have been expected based on the effects of section control on speed should be considered. The occurrence of crash migration or spillover effects is discussed in the last part of this section. 5.1. Exceptionally high/low crash numbers at two sites One of the sites had an exceptionally small number of injury crashes in the before period (Finstad). The estimated reduction of injury crashes is therefore larger when this site is omitted than when it is not omitted. A more detailed look at the road characteristics at this site indicates that it may rather be the predicted number of injury crashes that is exceptionally high. The predicted number of injury crashes per million vehicle kilometers is between 0.04 and 0.12 at other sites and 0.31 at Finstad, which is due to the combination of five at-grade junctions, three vertical grades, a 90 km/h speed limit and a road category that normally is associated with relatively high numbers of KSI. These factors contribute to high predicted crash numbers according to the crash prediction model. However, in reality the vertical grades are not severe, the at-grade junctions have only small proportions of secondary road volumes and the general road standard is quite high compared to other roads of the same category. Without Finstad, the difference between observed and predicted crash numbers in the before period is considerably reduced, and almost eliminated on open roads. Thus, for open roads the result for injury

Table 8 Effects of section control on speed (Ragnøy, 2011, 2013).

Barkald Rosten Dørdal Eiksund S Eiksund N/S Hell N Tromsøysund S Tromsøysund N All sites Open roads Tunnels

Speed before (km/h)

Speed after (km/h)

Effects on speed km/h %

88.5 89.4 76.7 84.4 79.3 77.9 80.3 79.6

78.3 80.6 74.0 75.2 74.8 75.3 73.6 73.7


10.2
8.8
2.7
9.2
4.5
2.6
6.7
5.9


12%
10%
4%
11%
6%
3%
8%
7%
6%
6%
5%

Crash effects, based on effect of speed Injury crashes KSI


9%
9%
9%


19%
20%
19%

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

crashes with Finstad omitted may be more representative of the effects of section control than the result for all sites. Another site had an exceptionally high number of KSI in the before period (Eiksund S which is located in an undersea tunnel). This can be regarded as a mere chance effect. All five KSI were fatalities in one crash (a head-on collision with a subsequent fire), and section control would have been installed at this site even without the crash. The contribution of this site to the effect of RTM must therefore be regarded as real and omitting this site (or reducing the number of KSI in the before period at this site for example to one) has no substantial influence on the results from the EB-evaluation. Thus, this site should not be omitted from the overall results or for the results for tunnels. The results from the BA study without control for RTM however are quite different depending on whether or not Eiksund S is omitted and likely to be strongly affected by RTM when Eiksund S is not omitted. 5.2. Overestimated effects of RTM in tunnels A comparison between the results from the BA study with comparison group (without control for RTM) and the EB evaluation (with control for RTM) indicates that there are regression effects in tunnels, both for injury crashes and KSI, and on open roads for KSI, but not for injury crashes. The differences between the results with and without control for RTM, and thus the regression effects, are greater for KSI than for injury crashes and greater in tunnels than on open roads. This is consistent with the finding that observed numbers of injury crashes exceeded the predicted numbers overall and in tunnels, but not on open roads, and that the observed numbers of KSI exceed the predicted numbers on open roads and in tunnels, more in tunnels than on open roads. However, the effect of RTM in tunnels may be overestimated. Firstly, high crash numbers were a criterion for installing section control on open roads but not in most of the tunnels. The large differences between predicted and observed crash numbers in tunnels are therefore contrary to expectation. Secondly, the predicted crash numbers in tunnels may be underestimated. Section control in most tunnels is located on steep downhill sections (most tunnels are undersea tunnels). Undersea tunnels are a predictor variable in the crash models that were used for calculating predicted crash numbers, but not the degree, length or direction of vertical grade. Grades in tunnels were in other studies found to contribute considerably to increasing crash numbers, especially to severe crashes (SWOV, 2011). The predicted crash numbers in tunnels may therefore be underestimated. The regression effects would then have been overestimated and the results from the EB-evaluation in tunnels would be underestimated. 5.3. Uni- vs. bidirectional section control Section control should be expected to have greater effects at sites with bidirectional section control than at sites with unidirectional section control. However, all except one of the unidirectional section control sites are in tunnels and all except one of the tunnel sites have unidirectional section control. Because of the confounding effect between location (tunnel vs. open road) and uni- vs. bidirectional section control and the small number of sites it was not possible to investigate the effects of uni- vs. bidirectional enforcement while controlling for other factors such as location. The unidirectional section control sites in tunnels have in common that all except one are on downhill segments in undersea tunnels. Two of them are in one of the tubes of a two-tube tunnel with traffic only in one direction. At these sites, all traffic can be assumed to be affected by section control. At the other unidirectional tunnel

177

sites, theoretically only half of all traffic can be assumed to be directly affected by section control. However, high speed is mostly a problem in the downhill direction (which is why section control only was installed in the downhill direction) and it is unlikely that speed above the speed limit is a contributing factor to many crashes in which a vehicle driving uphill has caused the crash. Section control can therefore not be expected to have equally large effects in the uphill direction as in the downhill direction. The estimated effects of section control can be interpreted as effects at bidirectional sites on open roads and at unidirectional tunnel sites. It cannot be concluded that section control on open roads will have the same effect when section control is installed in only one direction than when it is installed in both directions. 5.4. Crash vs. speed effects The effects that were found on injury crashes and KSI are considerably greater, even with control for RTM, than those effects that would have been expected based on the effects on speed and the general relationship between speed and crashes. This might indicate that the estimated effects on crashes are overestimated (there are however no specific factors that indicate that crash effects may be overestimated, only some that indicate that they even may be underestimated, see above), that the speed reductions for some reason are underestimated (for example because short term effects were measured while speed is reduced more in the long term), that the relationship between speed for some reason is stronger at section control sites than elsewhere, or that other factors than speed reductions have contributed to the crash effects. 5.5. Spillover effects on downstream sections Injury crashes were significantly reduced on downstream sections and even if the number of KSI has increased according to the EB evaluation, the result for KSI is hardly interpretable. Without control for RTM, the results shown a decrease of the number of KSI on downstream sections. Moreover, there was only one crash with KSI in the after period on downstream sections and the result is highly sensitive to the number of KSI in this crash. Had fewer people been severely injured in this crash, the result would have been a crash reduction that even might have been statistically significant. It is therefore concluded that the results do not indicate that concerns about crash migration are justified. On the contrary, spillover effects are more likely to occur. This is in line with result from other studies (ARRB, 2005; Chen et al., 2002). These did not find any evidence for crash migration either. For injury crashes the results indicate that the effect of section control is even greater on downstream sections than at the section control sites where no significant effect was found. However, the relatively small number of injury crashes makes it difficult to compare the size of the effect between section control sites and downstream sections. Speed data that might confirm or contradict this finding are not available for downstream sections. 6. Conclusions Section control was found to reduce both injury crashes and the number of KSI. When controlling for RTM the number of KSI was found to be statistically significantly reduced by 49%. For injury crashes a non-significant reduction by 12% was found. The results refer to sites with no fixed camera enforcement in the before period. Some of the sites had speed cameras in the before period, but the effects of these are statistically controlled for. For KSI no substantial differences were found between the effects of section control in tunnels and on open roads. When a site with exceptionally many KSI in the before period is omitted, the

178

A. Høye / Accident Analysis and Prevention 74 (2015) 169–178

estimated reduction of KSI becomes somewhat smaller in tunnels than on open roads, but when taking into account that regression effects most likely are overestimated in tunnels, and that the results from the EB evaluation therefore may be underestimated, the effects are most likely about the same in tunnels and on open roads. Since the confidence intervals are far larger than the difference between the effects on KSI in tunnels and on open roads it cannot be concluded that section control has different effects on KSI in tunnels and on open roads. The results for injury crashes indicate that section control has greater effects in tunnels than on open roads. In tunnels injury crashes were found to be reduced by 17%, and may be reduced by up to 25% when one assumes that the regression effects in tunnels are overestimated. On open roads, injury crashes were found to be reduced by 12%, or by 21% if one omits the site with exceptionally few (or unrealistically many predicted) injury crashes in the before period. The differences between the effects in tunnels and on open roads are however small in comparison to the large confidence intervals. The sizes of the effects that were found for injury crashes and KSI should be interpreted with some caution because only a limited number of crashes has been available for the evaluation and some of the results are quite sensitive to the outcomes of individual crashes. It is for example unclear why section control should have a greater effect on injury crashes on downstream sections than at the section control sites. On the other hand, the results are consistent with the assumption that section control has greater effects on more serious crashes, and the crash effects are consistent with the findings from speed measurements. Section control on open roads is at most sites installed in both directions. At most tunnel sites it is installed in only one direction. However, it cannot be concluded that the crash effects tunnels would be greater if section control were installed in both directions because the section control sites are steep downhill grades where it is unlikely that high speed is a contributing factor to many crashes in the opposite direction. The results of an evaluation of the effects of section control downstream of the section control sites indicate that spillover effects occur, i.e. that crashes are reduced even on downstream sections. This is contrary to the assumption that section control may cause crash migration because of drivers speeding up after having passed a section control site. Acknowledgement The Norwegian Public Roads Administration supported this work financially and provided the data for the study.

References ARRB, Evaluation of the fixed digital speed camera program in NSW. Report RC 2416 ARRB Consulting, 2005. Brassøe, B., Johansen, J.W., Madsen, J.C.O., Lahrmann, H., 2011. Sikkerhedsmæssig effekt af strækningshastighedskontrol i Storbritannien. Trafikdage på Aalborg Universitet. Broughton, P.S., Hutchings, C., Stone, D., Walker, L., 2012. Effectiveness of average speed cameras on the reduction of road casualties: Analysis of the A77 in Scotland. In: Dorn, L. (Ed.), Driver Behaviour and Training, Vol. V. Ashgate Publishing, Aldershot. Cameron, M.H., Elvik, R., 2010. Nilsson's power model connecting speed and road trauma: applicability by road type and alternative models for urban roads. Accid. Anal. Prev. 42 (6), 1908–1915. Chen, G., Meckle, W., Wilson, J., 2002. Speed and safety effect of photo radar enforcement on a highway corridor in British Columbia. Accid. Anal. Prev. 34 (2), 129–138. Decina, L.E., Thomas, L., Srinivasan, R., Staplin, L., 2007. Automated enforcement: a compendium of worldwide evaluations of results. TransAnalytics, LLC. 1722 Sumneytown Pike, Kulpsville, PA. Elvik, R., 2008. The predictive validity of empirical Bayes estimates of road safety. Accid. Anal. Prev. 40, 1964–1969. Elvik, R., 2009. The power model of the relationship between speed and road safety. TØI-report 1034. Oslo: Institute of Transport Economics. Elvik, R. 2012. Fartsgrenser (Speed limits). In: Trafikksikkerhetshåndboken (Handbook of Road Safety Measures). http//tsh.toi.no. Hauer, E., 1997. Observational Before–After Studies in Road Safety. Pergamon Press, Elsevier Science Ltd., Oxford, UK. Hauer, E., Harwood, D.W., Council, F.M., Griffith, M.S., 2002. Estimating safety by the empirical Bayes method: a tutorial. Trans. Res. Rec. 1784, 126–131. Høye, A., 2014a. Utvikling av ulykkesmodeller for ulykker på riks- og fylkesvegnettet i Norge (Development of crash prediction models for national and county roads in Norway). TØI Report 1323/2014. Oslo: Institute of Transport Economics. Høye, A., 2014b. Evaluering av effekt på ulykker ved bruk av streknings-ATK (Evaluation of the crash effects of section control). TØI Report (in press). Oslo: Institute of Transport Economics. Høye, A., 2014c. Automatisk trafikkontroll (Speed cameras and section control). In: Trafikksikkerhetshåndboken (Handbook of Road Safety Measures). http//tsh. toi.no. Montella, A., Persaud, B., D'Apuzzo, M., Imbriani, L.L., 2012. Safety evaluation of automated section speed enforcement system. Trans. Res. Rec. 2281, 16–25. Mountain, L.J., Hirst, W.M., Maher, M.J., 2004. Costing lives or saving lives: a detailed evaluation of the impact of speed cameras. Traffic, Eng. Control 45 (8), 280–287. Persaud, B., Council, F.M., Lyon, C., Eccles, K., Griffith, M., 2005. Multijurisdictional safety evaluation of red light cameras. Trans. Res. Rec. 1922, 29–37. Ragnøy, A., 2011. Streknings-ATK (Section control). VD Report Nr. 1. Norwegian Public Roads Administration. Ragnøy, A., 2013. Streknings-ATK i tunnel. Målt effekt på kjørefart, beregnet effekt på ulykker (Section control in tunnels. Effects on speed and estimated effect on crashes). Report Nr. 142. Norwegian Public Roads Administration. Stefan, C., Winkelbauer, M., 2005. Section control – automatic speed enforcement in the Kaisermühlen tunnel (Vienna, a22 motorway). Rosebud WP4Case Report. Wien: Kuratorium für Verkehrssicherheit. SWOV, 2011. The road safety of motorway tunnels. SWOV Fact Sheet. Leidschendam, The Netherlands. Tay, R., 2010. Speed cameras improving safety or raising revenue? J. Trans. Econ. Policy (JTEP) 44 (2), 247–257. Wilson, C., Willis, C., Hendrikz, J.K., Le Brocque, R., Bellamy, N., 2012. Speed cameras for the prevention of road traffic injuries and deaths. Cochrane Database of Systematic Reviews, 11(10 Article CD004607).