U.S. Department of Transportation
Federal Highway Administration
1200 New Jersey Avenue, SE
Washington, DC 20590
202-366-4000


Skip to content
Facebook iconYouTube iconTwitter iconFlickr iconLinkedInInstagram

Federal Highway Administration Research and Technology
Coordinating, Developing, and Delivering Highway Transportation Innovations

 
REPORT
This report is an archived publication and may contain dated technical, contact, and link information
Back to Publication List        
Publication Number:  FHWA-HRT-12-048    Date:  November 2013
Publication Number: FHWA-HRT-12-048
Date: November 2013

 

Pavement Marking Demonstration Projects: State of Alaska and State of Tennessee

CHAPTER 4. SAFETY EFFECTS OF WIDE EDGE LINES

INTRODUCTION

This chapter provides a description of the activities conducted in this study to better understand the safety effects of wider edge lines. First, a brief literature review is provided of previous research findings related to wider pavement markings, which were inconclusive in terms of demonstrating the safety effects of wider edge lines. Next, the data collection and data preparation activities conducted are described. Finally, the analyses and findings of the safety effect of wider edge lines are provided.

LITERATURE REVIEW

Many agencies are experimenting with enhanced pavement markings to reduce crashes and/or crash rates (i.e., adding markings to rural two-lane highways, adding wider edge lines, installing specially designed markings with relatively high retroreflectivity under wet conditions, etc.). Much of this emphasis has resulted from national programs such as AASHTO’s Implementing the AASHTO Strategic Highway Safety Plan, as described earlier.(6) Other factors, such as increased emphasis on accommodating older drivers, have also inspired agencies to evaluate their marking programs.

Studies have found that the use of markings plays a role in the reduction of specific crash types under certain conditions.(51-53) Run-off-road and opposite direction crashes are generally overrepresented on U.S. highways, especially on horizontal curves and at night when fatal crashes are three to four times more likely to occur. In addition, due to visual and cognitive deficiencies, older and impaired drivers are especially susceptible to these types of crashes. Therefore, crash types that are most likely affected by added or enhanced markings (added width or greater retroreflectivity) are run-off-road and opposite-direction crashes that occur at night, occur on curves, and involve drivers with reduced visual or cognitive capabilities (e.g., older drivers or impaired drivers).

Before-after crash studies conducted in Virginia and New Mexico in 1987 and 1988 suggest that wider lines have no safety benefit in terms of reducing crashes.(51,52) However, these studies were hampered by insufficient data and lack of experimental control.

DATA COLLECTION AND PREPARATION

This section summarizes the safety analysis efforts associated with various pavement marking widths and provides a general description of the data collection approach. The focus is RTLTW highways. The results of three types of analyses are then presented. The first analysis is a cross-sectional safety comparison of rural two-lane segments with 5-inch edge lines to similar segments with 4-inch edge lines. The second and the third analyses are an empirical Bayes (EB) before-after analysis and an interrupted time series analysis of rural two-lane segments on which the edge line width was changed from 4 to 6 inches.

Identifying Available Data

An electronic survey was distributed to identify States that have wider pavement markings (wider than 4 inches) on all or some of their State-owned highways. It was sent through several different media including the following:

Several rounds of follow-up telephone calls were made to identify States with previous or current wider line experience. State traffic engineers, district traffic engineers, maintenance engineers, and staff from other safety-related agency branches were contacted to determine whether the following factors applied:

The convergence of affirmative answers in all four areas was rare. Required data were most readily available in Illinois, Kansas, and Michigan.

Illinois Data Collection and Preparation

Illinois has varying pavement marking practices across its nine districts. The minimum line width in district 6 is 5 inches. This includes edge lines on both sides of the traveled way, skip lines, and other types of centerline markings. In district 3, edge lines and solid yellow centerline markings on two-lane highways are 4 inches wide, while skip lines on multilane highways and in two-lane highway passing zones are 6 inches wide. The pavement marking practices date back 15+ years before the availability of reliable crash and roadway data. Accordingly, a cross-sectional analysis approach was developed using current crash, traffic, and roadway data.

Illinois is a participating State in the Highway Safety Information System database (HSIS).(54) HSIS is a multi-State database managed by the University of North Carolina Highway Safety Research Center and Lendis Corporation under contract with FHWA. Participating HSIS States were selected based on their data quality and the ability to merge electronically coded crash-related and highway infrastructure-related files. The HSIS database is often the first data alternative for highway safety research with national sponsorship and geometric design components, including research efforts associated with production of Highway Safety Manual and SafetyAnalyst.(55,56)

Illinois crash and roadway inventory files for districts 6 and 3 from 2001 through 2008 were obtained from HSIS. Crashes were located by county, route number, and milepost. Roadway segments were defined by county, route number, begin milepost, and end milepost. Crashes were assigned to appropriate roadway segments and counted using a variation of SAS® code provided by the HSIS lab manager as well as with structured query language code written by the Texas Transportation Institute (TTI). Over 115 different crash type variations were originally counted. After preliminary model estimation runs and research team decisions were conducted related to the most relevant crash counts for this analysis, the number of crash types was reduced to the following:

Roadway segments and associated crash counts for rural two-lane highways, which are the focus of this report, were identified using area type and roadway classification indicators. Rural two-lane segments coded with presence of traffic signals, stop signs, or yield signs were deleted from the database to minimize the influence of intersection presence on the analysis. Additional segments coded as having atypical rural two-lane highway features (e.g., medians, auxiliary lanes, etc.) were also eliminated. Finally, segments that showed any change in physical features during the observation period were deleted to minimize the influence of any major reconstruction project on the analysis results. The number of segments included in the analysis decreased as the number of observation years increased as a result of this criterion.

A large number of segments were redefined in 2007 and 2008. The rural two-lane dataset for Illinois for 2001 through 2006 included 6,531 segments (1,733 mi): 5,343 segments (1,446 mi) with 4-inch edge lines and centerlines and 1,188 segments (287 mi) with 5-inch edge lines and centerlines. From 2001 through 2008, the number of segments reduced to 3,214 segments (643 mi): 2,572 segments (520 mi) with 4-inch edge lines and centerlines and 642 segments (123 mi) with 5-inch edge lines and centerlines. The 2001 through 2006 rural two-lane highway database was used for the final analysis to preserve the larger sample of segments. Table 29 and table 30 summarize the descriptive statistics for the primary segment variables considered in the analysis.

Table 29. Descriptive statistics for continuous Illinois rural two-lane highway segment variables.


Segment Variable

5,343 Segments (1,446 mi)
with 4-Inch-Wide Markings

1,188 Segments (287 mi)
with 5-Inch-Wide Markings

Minimum

Maximum

Average

Minimum

Maximum

Average

Length (mi)

0.01

5.45

0.27

0.01

2.51

0.24

ADT (vehicles)

100

25,900

3,316

100

11,100

2,140

Daily commercial traffic (trucks)

0

4,500

379

0

1,100

281

Lane width (ft)

8

16

11.7

9

16

11.6

Shoulder width (ft)

0

14

6.5

0

12

6.0

Paved shoulder
width (ft)

0

14

3.8

0

12

4.3

Table 30. Descriptive statistics for categorical Illinois rural two-lane highway segment variables.


Segment Variable

5,343 Segments (1,446 mi)
with 4-Inch-Wide Markings

1,188 Segments (287 mi)
with 5-Inch-Wide Markings

Frequency

Percent

Frequency

Percent

Posted speed = 25 mi/h

6

0.1

4

0.3

Posted speed = 30 mi/h

180

3.4

62

5.2

Posted speed = 35 mi/h

301

5.6

73

6.1

Posted speed = 40 mi/h

287

5.4

45

3.8

Posted speed = 45 mi/h

302

5.7

86

7.2

Posted speed = 50 mi/h

164

3.1

21

1.8

Posted speed = 55 mi/h

4,103

76.8

897

75.5

Presence of horizontal curve sharper than 2.5 degrees

962

18.0

140

11.8

Kansas Data Collection and Preparation

Kansas began installing 6-inch edge lines on all State-owned roads in July 2005. Implementation was not immediate but was accomplished during normal construction and maintenance activities. An email to Kansas district engineers, maintenance engineers, and maintenance paint crews dated July 7, 2005, contains instructions to begin painting edge lines 6 inches wide on all projects beginning as soon as provisions could be made to accommodate 6-inch tips on guns and glass bead shrouds.

Changes from 4- to 6-inch edge lines in Kansas occurred primarily from 2005 through 2009. Data related to the timing and locations of these changes are available for districts 2 and 6. There were some segments in these districts where edge lines were not changed from 4 to 6 inches until 2008 or later. There are a few remaining segments in Kansas that still have 4-inch edge lines. The structure of the data allowed an EB analysis. A majority of the available data in districts 2 and 6 were for rural two-lane roadways. This facility type is the focus of the analysis in Kansas.

Crash and roadway data were obtained directly from Kansas Department of Transportation’s (KDOT) Bureau of Transportation Planning. Pavement marking data were obtained from maintenance engineers in districts 2 and 6. While data were available for 2000 through 2009, data for 2001 through 2007 were ultimately used for analysis due to incomplete roadway data for 2000 and incomplete crash data for 2009, as well as to increase the number of segments in the reference group (i.e., include road segments where edge lines were not changed until 2008 and later in the reference group). There was a group of 718 segments where 6-inch edge lines were implemented but the implementation year was unknown. Analysis was conducted with and without these segments, and a conservative estimate of implementation year was made for the analysis.

Crashes were located by county, route number, and county milepost, while roadway segments were located by county, route number, and begin and end county milepost. Crashes were assigned to appropriate roadway segments and counted using a variation of the SAS® and structured query language codes used for the Illinois data. The following types of crashes were counted for analysis:

The roadway segment definitions (i.e., begin and end county mileposts) were dependent on the procedure used to query the data. Segments were often defined differently from year to year even if the features describing the segment (e.g., lane width, shoulder width, etc.) did not change. Several variables, which differed from project to project, were used to define the pavement marking conversions from 4 to 6 inches Available information and notations were different between districts 2 and 6. The following sections summarize a manual roadway segment data coding process that included a variable for pavement marking width. It begins with a discussion of district 6 data, which was more detailed than district 2 and was used to develop the data coding procedure. Adjustments to the coding procedure were then made to accommodate the less detailed district 2 data.

Pavement Marking Data

District 6 provided information on the date and locations of installations (with locations defined in a variety of ways) from July 2005 (when the wider line policy was implemented) through August 2007. Table 31 shows an example of these data in raw form.

Table 31. Example of Kansas district 6 pavement marking data from April to June 2006.

Date

Route Striped

Paint Used (gal)

Beads Used (lb)

Miles Painted (mi)

4/17

K-23 Meade Co. Meade north city limit to reference post (R.P). 34.0 (both sides)

545

6,537

14.0

4/19

K-23 Meade Co. R.P. 34.0 to R.P. 37.3 (both sides)

 

 

 

4/19

K-23 Meade Co. R.P. 38.3 to R.P. 41.5 (both sides)

314

3,929

9.8

4/21

K-23 Gray/Meade Co. Jct US 56/K-23 to R.P. 38.3 (west side) and R.P. 41.5 to Jct US 56/K-23
(east side)

743

8,856

 

4/26

K-23 Gray Co. Cimarron south city limit to Jct
US 56/K-23 (both sides)

793

9,649

24.0

4/27

US 283 Clark Co. Kansas/Okla. State line to South Jct US-283/US-160 (east side)

 

 

13.6

4/27

US 283 Clark Co. South Jct US 160/US 283 to
R.P. 9.5 (west side)

554

6,834

4.1

5/2

US 283 Clark Co. R.P. 9.5 to Kansas/Okla. State line (west side)

298

3,684

9.5

5/8

K-96 Scott Co. Scott/Lane Co. line to R.P. 67.0 (both sides)

466

5,825

14.0

5/12

K-96 Scott Co. R.P. 67.0 to Scott City east city limit (both sides)-includes lane lines in Scott City

631

7,572

8.0

5/16

US 83 Finney Co. R.P. 70.6 (end of concrete) to R.P. 85.0 (both sides)

774

9,506

31.2

5/17

US 83 Finney/Scott Co. R.P. 85.0 to Scott City south city limit (east side)

 

 

23.4

5/17

US 83 Scott Co. Scott City south city limit to
R.P. 101.4 (west side)

811

10,154

7.0

5/18

US 83 Scott/Finney Co. R.P. 101.4 to R.P. 85.0 (west side)

572

6,951

16.4

5/19

K-95 Scott Co. North Jct US 83/K-95 to south
Jct US 83/K-95 (west side)

228

2,850

6.5

5/22

K-95 Scott Co. South Jct US-83/K-95 to north
Jct US 83/K-95 (east side)

214

2,645

6.5

6/6

US 50 Kearny Co. Finney/Kearny Co. line to Lakin east city limit (north side)

356

4,450

9.5

6/7

US 50 Kearny/Hamilton Co. Lakin west city limit to Syracuse east city limit (north side)

873

10,645

26.3

6/13

US 50 Hamilton Co. Syracuse east city limit to Kearny/Hamilton Co. line (south side)

180

5,376

11.3

6/14

US 50 Kearny Co. Hamilton/Kearny Co. line to Lakin west city limit (south side)

283

6,344

18.6

6/15

US 50 Kearny Co. Lakin east city limit to Finney/Kearny Co. line (south side)

180

4,283

9.4

Note: Blank cells indicate that data were unavailable.

The research team developed a method to convert the information in the "Route Striped" column to a beginning and ending State milepost. Route numbers and county names were always provided, but the extent of a striping project on any given day and route was defined by one or more of the following features:

A few examples of converting the "Route Striped" descriptions to beginning and ending State mileposts are provided in the following sections.

Example 1. Mileposts directly provided. The State mileposts are provided in the "Route Striped" column of table 31 in this case. For example, the western side of US 83 was striped on May 18, 2006, from R.P. 101.4 to R.P. 85.0. The road sections of interest were then identified in the Kansas roadway files (see table 32 for an example excerpt from the road file where segments were painted with wider lines on May 18, 2006). Several other Kansas variables are available in the road file. Table 32 is for illustration purposes only.

Table 32. Excerpt from Kansas road file for US 83.

Longitudinal Reference System (LRS)

Begin
ST_MP

End
ST_MP

AADT

Heavy
Commercial

Shoulder Width
(inches)

028U0008300S0

79.107

86.107

3,410

1,250

3

028U0008300S0

86.107

93.328

3,230

1,250

3

086U0008300S0

93.328

97.628

3,190

1,260

3

086U0008300S0

97.628

100.628

3,510

1,260

3

086U0008300S0

100.628

103.628

3,700

1,270

3

ST_MP = State milepost.

Example 2. Striping project defined from State line to road junction. In this example, the "Route Striped" column provides the extents of the striping project from the State line to a road junction. The eastern side of US 283 in Clark County was striped from the Kansas/Oklahoma State line to the south junction of US 283 and US 160 on April 27, 2006 (see table 31). County maps, available through the KDOT Web site, were used to convert this type of information to beginning and ending State mileposts. For this example, the Kansas Clark County map was used (see figure 12). A screenshot of a portion of the Clark County map shows that US 283 is a northbound-southbound route from the Kansas/Oklahoma State line through Clark County. The State milepost is zero at the southern State border and increases from south to north.

The south junction of US 283 and US 160 is shown in the top part of the screenshot. Road segments on the Kansas county maps contain asterisk (*) symbols. An approximate distance between these asterisks is provided at the midpoint of each segment. The example in figure 12 shows that the distance between the asterisk at the Kansas/Oklahoma border and the asterisk at the south junction of US 283 and US 160 is approximately 13.8 mi. (Green circles have been added to the screenshot to help identify the asterisks at the beginning and end of the segment and the distance indication.)

This figure shows a screenshot of a map of Clark County, KS. Green circles along route US 283 have been added to the screenshot to help identify the asterisk symbols at the beginning and end of the segment used in the evaluation and to identify the distance indication for the segment (13.6 mi).
Figure 12. Screenshot. Map of Clark County, KS

Next, the respective road segments in the Kansas road files were identified. Table 33 provides an excerpt from the road file data for US 283 in Clark County. The table shows that the junction of US 283 and US 160 occurs at milepost 13.579. Therefore, the road segments represented in the first nine rows were striped with wider lines on April 27, 2006.

Table 33. Excerpt from Kansas road file for US 283.


LRS

Begin
ST_MP

End
ST_MP

Begin
CO_MP

End
CO_MP

AADT

Heavy
Commercial

Shoulder
Width (inches)

013U0028300S0

0

2.034

0

2.034

710

285

3

013U0028300S0

2.034

2.049

2.034

2.049

1,020

225

3

013U0028300S0

2.049

2.336

2.049

2.336

1,020

225

0

013U0028300S0

2.336

2.683

2.336

2.683

1,110

245

0

013U0028300S0

2.683

3.046

2.683

3.046

940

255

2.7

013U0028300S0

3.046

3.35

3.046

3.35

940

255

3

013U0028300S0

3.35

10.557

3.35

10.557

710

290

3

013U0028300S0

10.557

13.355

10.557

13.355

715

300

3

013U0028300S0

13.355

13.579

13.355

13.579

715

300

1.8

013U0028300S0

20.16

26.16

20.16

26.16

725

275

3

013U0028300S0

26.16

27.16

26.16

27.16

875

260

3

013U0028300S0

27.16

31.285

27.16

31.285

865

260

3

013U0028300S0

31.285

31.388

31.285

31.388

1,530

275

0

013U0028300S0

31.388

31.568

31.388

31.568

1,530

275

0

013U0028300S0

31.568

31.672

31.568

31.672

1,790

265

0

013U0028300S0

31.672

31.734

31.672

31.734

2,480

265

3

013U0028300S0

31.734

31.927

31.734

31.927

2,010

560

3

013U0028300S0

31.927

33.718

31.927

33.718

2,010

560

3

ST_MP = State milepost.
CO_MP = County milepost.
Note: Shading indicates that road segments were striped with wider lines on April 27, 2006.

Example 3. Striping project defined from city limit to county line. The final example involves a "Route Striped" description that provides the extents of the striping project from a city limit to a county line. The north side of US 50 in Kearny County from the Finney-Kearny County line to the Lakin east city limit was restriped on June 6, 2006 (see table 31). The Kearny County map shows that US 50 is an eastbound-westbound route running the entire length Kearny County. Figure 13 shows a screenshot of the segment of interest from Lakin City to the Finney/Kearny County line. (Green circles have been added to the screen shot to help identify the asterisk symbols at the beginning and end of the segment and the distance indication.) Table 34 provides an excerpt from the road file data for US 50 in Kearny County. Segments were painted with wider lines on June 6, 2006, which is evident by the shading of the last seven rows.

This figure shows a screenshot of a map of Kearny County, KS. Green circles along route US 50 have been added to the screenshot to help identify the asterisk symbols at the beginning and end of the segment used in the evaluation and to identify the distance indication for the segment (9.9 mi).
Figure 13. Screenshot. Map of Kearny County, KS

Table 34. Excerpt from Kansas road file for US 50.


LRS

Begin
ST_MP

End
ST_MP

Begin
CO_MP

End
CO_MP

AADT

Heavy Commercial

Shoulder Width (inches)

Speed Limit (mi/h)

047U0005000S0

28.498

29.776

0

1.278

1,860

740

3

65

047U0005000S0

29.776

30.497

1.278

1.999

1,860

740

3

65

047U0005000S0

30.497

30.668

1.999

2.17

1,860

740

3

65

047U0005000S0

30.668

32.035

2.17

3.537

1,860

740

3

65

047U0005000S0

32.035

34.875

3.537

6.377

1,860

740

3

65

047U0005000S0

34.875

35.498

6.377

7

1,860

740

1.8

65

047U0005000S0

35.498

35.729

7

7.231

1,930

740

1.8

65

047U0005000S0

35.729

37.482

7.231

8.984

1,930

740

3

65

047U0005000S0

37.482

37.776

8.984

9.278

2,030

750

3

65

047U0005000S0

37.776

39.196

9.278

10.698

2,030

750

3

65

047U0005000S0

39.196

42.163

10.698

13.665

2,030

750

3

65

047U0005000S0

42.163

42.498

13.665

14

2,030

750

1.8

65

047U0005000S0

42.498

42.784

14

14.286

2,470

770

1.8

65

047U0005000S0

42.784

43.461

14.286

14.963

2,470

770

3

65

047U0005000S0

43.461

43.818

14.963

15.32

2,470

770

0

40

047U0005000S0

43.818

43.965

15.32

15.467

5,730

795

0

40

047U0005000S0

43.965

44.104

15.467

15.606

6,160

940

0

40

047U0005000S0

44.104

44.411

15.606

15.913

4,010

970

0

40

047U0005000S0

44.411

48.98

15.913

20.482

4,010

970

3

65

047U0005000S0

48.98

49.08

20.482

20.582

4,190

970

3

65

047U0005000S0

49.08

49.78

20.582

21.282

4,190

970

3

65

047U0005000S0

49.78

51.276

21.282

22.778

4,190

970

3

65

047U0005000S0

51.276

51.661

22.778

23.163

4,190

970

3

55

047U0005000S0

51.661

51.665

23.163

23.167

4,190

970

3

65

047U0005000S0

51.665

53.865

23.167

25.367

4,320

1,040

3

65

ST_MP = State milepost.
CO_MP = County milepost.
Note: Shading indicates that these road segments were striped with wider lines on June 6, 2006.

County mileposts are zero at western and southern county lines and increase from west to east and from south to north. The final row of table 34 is the last segment of US 50 in Kearny County. US 50 crosses into Finney County at county milepost 25.367 (State milepost 53.865). Asterisks and the respective segment length in figure 13 indicate that the distance from the US 50/K25 junction to the Finney-Kearny County line is 9.9 mi. Therefore, the US 50/K25 junction is at county milepost 15.467 (25.367 - 9.9) and State milepost 43.965 (53.865 - 9.9). The posted speed limit changes from 40 to 65 mi/h at State milepost 44.411. This is probably the Lakin east city limit. This was verified using Google Maps® and Google Earth® to measure the distance from the US 50/K25 junction to the Lakin east city limit (see figure 14). This distance is 0.44 mi, which means the Lakin east city limit is at county milepost 15.907 (15.467 + 0.44) and State milepost 44.405 (43.965 + 0.44). These numbers are very close to where the speed limit changes from 40 to 65 mi/h in table 34. Therefore, it was concluded that the road segments represented in the last seven rows of table 34 were striped with wider lines on June 6, 2006.

This figure shows a satellite photo of the US 50/K25 junction in Kansas. A 0.44-mi stretch of US 50 is highlighted in yellow.
Figure 14. Photo. Measurement from US 50/K25 junction to Lakin east city limit

The pavement marking records for Kansas district 2 were not as specific as those for district 6. The major features referenced by the maintenance engineer for locating the termini of the pavement marking jobs were generally the same and included State mileposts, city limits, junctions, county lines, and State lines. The method used by the research team to identify the mileposts associated with these features was identical to that discussed in the preceding section for district 6. The pavement marking records for district 2 did not provide county names where the marking installation took place. The information from district 2 also did not include the date of the striping job. The records only showed the job timing by year.

Table 35 provides an example of the district 2 striping data. Data collection involved additional searching through county maps to find the locations of interest. The yearly data also limited the ultimate level of data aggregation used for data analysis.

Table 35. Example of district 2 striping data.


Route

Route Description

Miles

Center Line

Edge Line

White

Yellow

Beads

K-15

K-15/K-18 W Jct to the DK-MN line

33

33

66

1,452

726

11,616

K-15

K-18 E Jct to the SCL of Clay County

23.5

23.5

0

517

2,068

K-15

K-15 W to K-15 E

4

4

8

176

88

1,408

K-15

US-36 to the KS/NE State line

0

0

0

K-18

OT-DK to the K-15 Jct

9

9

18

396

198

3,168

K-18

Jct K15 to the US-77 Jct

14

14

28

616

308

4,928

K-18

I-70 to Ogden (K-114)

4

4

8

176

88

1,408

K-43

I-70 to the K-4 Jct (Hope)

21

21

42

924

462

7,392

I-70

Abilene to milepost 303

18

18

36

792

396

6,336

I-70

Niles Road

6

6

12

264

132

2,112

K-115

Palmer Road

0.5

0.5

1

22

11

176

K-148

K-15/148 to the WS/RP line

17

17

34

748

374

5,984

K-157

Rock Springs 4-H Camp Road

4

4

8

176

88

1,408

K-197

Industry Road

2

2

4

88

44

704

K-189

Miltonville Road

1

1

2

44

22

352

K-206

Chapman Road

1

1

2

44

22

352

K-209

Woodbine

2.5

2.5

0

55

220

US 24

Clay W to K-189

16.5

16.5

33

726

363

5,808

US 36

Washington Road to the K-22 Jct

16

16

0

352

1,408

US 36

Washington Road to the multilane start

10

10

20

440

220

3,520

BUSS-40

Super 8 to end Business 40

2

2

4

88

44

704

Roadway Data

Two separate roadway databases, including a 6-inch stripe timing variable, were built for districts 6 and 2 and later combined for analysis. The databases spanned 2001 through 2007. 2001 was selected as the beginning of the observation period because data for several roadway segments were missing for 2000. 2007 was selected as the end of the observation period to increase the number of segments in the reference group (i.e., to include road segments where edge lines were not changed until 2008 and later in the reference group). The remainder of this section describes the procedure developed to build Kansas databases for districts 6 and 2 with roadway and pavement marking characteristics from 2001 through 2007. A similar procedure was applied to build both databases. Important differences are identified.

Data Screening:

The raw data files received from KDOT included statewide routes. The files for different years did not always have the same set of variables. Data screening procedures were developed to extract rural two-lane highway segments in districts 2 and 6 and remove unnecessary variables. This process simplified the data files and transformed them into an identical format with the same set of variables across years, a characteristic needed in order to merge the files into a single dataset.

The data files had several variables for segment location and functional classification. The variables used to extract the rural two-lane roadway segments included the following:

All two-lane rural highways were extracted using the following five steps:

  1. Filter RSE_COUNTY: Keep all roadway segments located in counties of the district of interest.
  2. Filter LNCL_DESCR: Keep all roadway segments defined as "2LU_Two Lane, Undivided."
  3. Filter FUNC_DESCR: Remove roadway segments described as urban.
  4. Filter LN1L_DESCR: Keep roadway segments with through lanes, left-turn lanes, right-turn lanes, and passing or creeper lanes (e.g., truck lanes), and remove roadway segments with parking lanes.
  5. Filter LN1R_DESCR: Follow step 4 except for the opposite direction of travel.

These steps, applied to data from both districts 2 and 6, produced seven data files (one for each year from 2001 through 2007) with rural two-lane highway segments in those districts. Each of these data files contained a set of 50 or more variables. Many of these were not applicable to rural two-lane highways. Only variables of interest were retained, including year, county, route number, beginning and ending mileposts (State and county), segment length, functional classification, daily traffic, daily truck traffic, speed limit, lane widths, and shoulder widths and types.

Defining Road Segments:

Segment definitions (i.e., beginning milepost and ending milepost) were nearly identical for years 2002 through 2006. Therefore, the 2002 data file was selected as a "base" file for segment definition. Roadway segments for other years where all relevant variable values had not changed from 2002 were redefined to match the 2002 segment definition. The dataset was initially structured so that one row equaled one segment observed for 1 month. The rows were then aggregated so that one row equaled one segment observed for 1 year due to the level of detail available in the pavement markings data for district 2. The dataset building process consisted of the following steps:

  1. Copy an entire row (i.e., segment) from the 2002 roadway data file to a "Combined Data" file and copy this same row to the next 11 rows (a total of 12 rows for 12 months). Create a column for "Month" and "Year" and fill in these columns with appropriate values.
  2. Locate the same segment in other data files (2001 and 2003-2007). If the beginning and ending mileposts of the segment match those in the 2002 data file, copy the entire row to the "Combined Data" file. Repeat this 11 more times as described in step 1 for a total of 12 rows representing 12 months of the year. If data from other years have different beginning and ending mileposts from those of 2002 (i.e., the base year), redefine the segment by combining short segments or splitting up longer segment to match the 2002 segment. Check all variables to make sure that variable values on the segments being combined are the same.

    In this study, the estimated AADT was different in some cases, and a weighted average AADT was computed. The result of this process was 84 observations (i.e., rows) for each road segment: one row per month over 7 years. The data were later aggregated, resulting in seven observations (i.e., rows) for each road segment: one row per year over 7 years.

  3. Create a variable, "Time_paint," to indicate the timing of the 4- to 6-inch edge line conversion. "Time_paint" equals zero if the row in the roadway dataset is before the 6-inch edge line conversion, equals 1 for the month and year when the 6-inch edge line is first painted on the segment, equals 2 if the row in the roadway dataset is after the 6-inch edge line conversion, and equals 3 if the timing of the conversion is unknown.

    There were 718 segments where 6-inch edge lines were implemented but the implementation year was unknown. For these segments, "Time_paint" equaled 3. Analysis was conducted without (analysis 1) and with (analysis 2) these 718 segments, and a conservative estimate of implementation year was made for analysis 2. During the course of several years, some segments were painted more than once. The timing of the first edge line conversion from 4 to 6 inches was coded as "Time_paint" equals 1, with all following rows equaling 2 (i.e., the after period). Table 36 and table 37 provide the descriptive statistics for analysis 1 and analysis 2, respectively.

Table 36. Descriptive statistics for Kansas rural two-lane highway segments (analysis 1).


Segment Variable

Treatment Group
1,615 Segments (1,165.3 mi)

Reference Group
261 Segments (158.1 mi)

Minimum

Maximum

Average

Minimum

Maximum

Average

Length (mi)

0.002

8.1

0.72

0.005

6.16

0.61

ADT (vehicles)

65

12,800

1,036

40

4,745

746

Daily commercial traffic (trucks)

3

1,790

217

5

540

148

Lane width (ft)

10

14

11.8

11

15

11.5

Shoulder width (ft)

1

10

4.7

0

10

4.1

Paved shoulder
width (ft)

0

10

1.4

0

8

0.7

Table 37. Descriptive statistics for Kansas rural two-lane highway segments (analysis 2).


Segment Variable

Treatment Group
2,333 Segments (1,909.9 mi)

Reference Group
261 Segments (158.1 mi)

Minimum

Maximum

Average

Minimum

Maximum

Average

Length (mi)

0.002

8.1

0.82

0.005

6.16

0.61

ADT (vehicles)

65

12,800

1,238

40

4,745

746

Daily commercial traffic (trucks)

3

2,260

293

5

540

148

Lane width (ft)

10

15

11.9

11

15

11.5

Shoulder width (ft)

0

12

5.4

0

10

4.1

Paved shoulder
width (ft)

0

12

2.0

0

8

0.7

Michigan Data Collection and Preparation

Michigan edge lines are currently 6 inches wide on all State-owned roadways (except those with curbs and gutters). The change was made from 4-inch edge lines in 2004. An MDOT pavement marking engineer estimated that 6-inch lines were installed on 95 percent of applicable mileage in 2004, with the remainder installed in early 2005. A before-after analysis was possible with the timing of the change. The widespread switch from 4- to 6-inch edge lines minimizes the concern of selection bias or regression to the mean. However, it does not allow a before-after analysis using comparison sites within the same State. The research team examined several comparison site alternatives but ultimately used an interrupted time series analysis.

Michigan crash data from 2001 through 2009 were obtained from the Michigan State Police Traffic Crash Reporting Unit. MDOT provided roadway inventory files for those years. Crashes were located by county, route number, physical reference number, and milepost, while roadway segments were defined by county, route number, physical reference number, beginning milepost, and ending milepost. Crashes were assigned to appropriate roadway segments and counted using SAS® and structured query language. Counts for 14 of the 15 crash types available for Illinois were also available for Michigan (crash type 15, total number of fixed object crashes, was not available).

Roadway segments and associated crash counts for rural two-lane highways were identified using an area type indicator and a variable for total number of through lanes. Similar data screening techniques and criteria as those employed for Illinois data were used for Michigan, including those for intersections, atypical rural two-lane highway features, and observed changes in physical features during the observation period. Analysis of two Michigan datasets was reported: (1) a dataset for years 2001 through 2007 with 253 segments (851.5 mi) and (2) a dataset for years 2001 through 2009 with 238 segments (787.8 mi). Each segment was observed for 3 years with 4-inch lines (2001 through 2003) and for 3-5 years with 6-inch lines (2005 through 2007 or 2005 through 2009, depending on the dataset). Table 38 and table 39 summarize descriptive statistics for the primary segment variables considered in the 2001 through 2007 analysis. The descriptive statistics are very similar for the 2001 through 2009 dataset and are not reported.

Table 38. Descriptive statistics for continuous Michigan segment variables.


Segment Variable

253 Segments (851.5 mi)
with 4-Inch Edge Lines for 3 Years (2001-2003)
and 6-Inch Edge Lines for 3 Years (2005-2007)

Minimum

Maximum

Average

Length (mi)

0.04

12.69

3.37

ADT (vehicles)

200

18,600

4,470

Daily commercial traffic (trucks)

20

2,200

350

Lane width (ft)

10

12

11.5

Shoulder width (ft)

3

12

8.1

Paved shoulder width (ft)

0

11

4.2

Table 39. Descriptive statistics for categorical Michigan segment variables.


Segment Variable

253 Segments (851.5 mi)
with 4-Inch Edge Lines for 3 Years (2001-2003)
and 6-Inch Edge Lines for 3 Years (2005-2007)

Frequency

Percent

Posted speed = 25 mi/h

5

2.0

Posted speed = 30 mi/h

1

0.4

Posted speed = 35 mi/h

4

1.6

Posted speed = 40 mi/h

3

1.2

Posted speed = 45 mi/h

10

4.0

Posted speed = 50 mi/h

4

1.6

Posted speed = 55 mi/h

226

89.3

Level terrain

165

65.2

Rolling terrain

88

34.8

WIDER LINE RETROSPECTIVE CRASH ANALYSES

This section focuses on the analysis of wider edge lines on rural two-lane highways (findings from analyses on multilane highways are provided in appendix D). The results of safety analyses for three States are presented (Illinois, Kansas, and Michigan). Three separate analyses were required due to unique characteristics of the data, including how, when, and the extent to which States made the transition to wider lines, as well as how long it took the States to complete the transition. The first analysis is a cross-sectional safety comparison of rural two-lane segments with 5-inch centerlines and edge lines to segments with 4-inch centerlines and edge lines for Illinois. The second analysis is an EB before-after analysis of rural two-lane segments in Kansas for which the edge line width was changed from 4 to 6 inches in multiple years. The third analysis is a piecewise regression analysis of interrupted time series design with the change from 4 to 6 inches in 2004 being treated as an intervention for Michigan data.

Analysis of Illinois Rural Two-Lane Roadway Crash Data

Illinois crash data from 2001 through 2006 were obtained from 6,531 segments, which roughly corresponded to 1,733 mi of rural two-lane highways. Out of the 6,531 segments, 5,343 segments (1,446 mi) have 4-inch edge lines and 4-inch centerlines and 1,188 segments (287 mi) have 5-inch edge lines and 5-inch centerlines. Crashes occurring at the segments with 4-inch edge lines were compared to crashes occurring at the segments with 5-inch edge lines using the cross-sectional data analysis. Only the non-intersection/interchange crashes were considered. Crashes occurring during the winter months (November through March) were excluded from the analysis to avoid any potential confounding by snow crashes.

During the course of data analysis, it was revealed that about 50 percent of total crashes (about 60 percent of PDO crashes, 60 percent of single-vehicle crashes, and 10 percent of F+I crashes) were animal collisions. While animal collisions were deemed to be irrelevant for assessing safety effects of wider edge lines, the proportion of animal collisions was significant. Therefore, researchers conducted cross-sectional analyses for two different Illinois datasets from 2001 through 2006, one with animal collisions included and the other with animal collisions excluded. Types of crashes analyzed included the following:

Table 40 summarizes the 2001 through 2006 Illinois crash datasets used for the analysis. The table shows the aggregated crash counts and crash rates computed as crashes per million vehicle miles of travel per year (non-winter month crash counts for 7 months were first multiplied by a factor of 12 divided by 7 before computing crash rates) for Illinois rural two-lane highways. It is categorized by edge line width for each of two datasets (dataset with animal collisions included and dataset without animal collisions).

Table 40. Summary of Illinois crash data for 2001-2006 used in the analysis.


Variable

Dataset With Animal Collisions

Dataset Without Animal Collisions

4 inches

5 inches

4 inches

5 inches

4 inches

5 inches

4 inches

5 inches

Number of segments

5,343

1,188

5,343

1,188

5,343

1,188

5,343

1,188

Total segment length (mi)

1,446.1

286.7

1,446.1

286.7

1,446.1

286.7

1,446.1

286.7

Mile-years

8,676.8

1,720.1

8,676.8

1,720.1

8,676.8

1,720.1

8,676.8

1,720.1

Average AADT

3,316.0

2,339.9

3,316.0

2,339.9

3,316.0

2,339.9

3,316.0

2,339.9

Crash Type

Crash Counts

Crash Rates

Crash Counts

Crash Rates

Total

6,135

957

1.00

1.12

3,397

342

0.55

0.40

F+I

1,595

169

0.26

0.20

1,451

137

0.24

0.16

PDO

4,540

788

0.74

0.92

1,946

205

0.32

0.24

Daytime

2,802

331

0.46

0.39

2,213

219

0.36

0.26

Nighttime

2,805

504

0.46

0.59

1,027

109

0.17

0.13

Daytime F+I

964

92

0.16

0.11

919

88

0.15

0.10

Nighttime F+I

555

69

0.09

0.08

466

46

0.08

0.05

Wet

666

78

0.11

0.09

464

45

0.08

0.05

Wet nighttime

297

36

0.05

0.04

155

16

0.03

0.02

Single-vehicle

4,669

818

0.76

0.95

1,942

203

0.32

0.24

Single-vehicle wet

519

69

0.08

0.08

317

36

0.05

0.04

Single-vehicle nighttime

2,581

485

0.42

0.57

810

90

0.13

0.11

Single-vehicle daytime F+I

1,025

122

0.17

0.14

884

90

0.14

0.11

Single-vehicle night F+I

455

60

0.07

0.07

368

37

0.06

0.04

Older driver
(≥ 55 years old)

1,280

195

0.21

0.23

706

74

0.12

0.09

Fixed object

1,127

133

0.18

0.16

1,127

133

0.18

0.16

The crash rates shown in table 40 might be useful if all of the segments included in the study were identical except for edge line width, segment length, and AADT and also if crashes increased linearly with AADT. However, the road segments were different not only in edge line width, segment length, and AADT, but also in other roadway characteristics such as lane width, shoulder width, presence of curves, etc. Additionally, the relationship between crashes and AADT was not necessarily linear. As a result, the effects of edge line width may not be estimated correctly by the differences in simple crash rates between 4- and 5-inch edge line segments.

In order to separate the effect of edge line width from other important roadway characteristics, the negative binomial regression models (or Poisson regression models when negative binomial regression models could not be fitted) were applied to these cross-sectional data. The general form of the expected number of crashes in a negative binomial regression model is given in figure 15.

Mu subscript i equals exp times open parenthesis beta subscript zero plus beta subscript l times X subscript 1i plus beta subscript 2 times X subscript 2i plus ellipsis plus beta subscript k times X subscript ki close parenthesis.
Figure 15. Equation. General form of the mean of negative binominal regression

Where μ is the expected number of crashes at segment i, X1i, ..., Xki are the covariates/predictors corresponding to roadway characteristics of segment i, and β0, β1, β2,..., βk are the regression coefficients. After exploring various negative binomial regression model forms with different predictors and interaction terms, the model including wider edge line (coded as "1" when edge line width = 5 inches and "0" when edge line width = 4 inches), lane width, shoulder width, log of AADT, presence of horizontal curve with degree of curve greater than 2.5 degrees (1 = present, 0 = not present), and log of segment length as predictors seemed to be most appropriate for these data. The horizontal curve indicator variable was created using the non-zero entries for horizontal curve beginning and ending mileposts contained in the HSIS road files. A comparison of the curve mileposts to the road segment mileposts indicated that the entire road segment with the curve presence indicator variable equal to 1 was located inside the boundaries of the horizontal curve. The natural logarithm of segment length may be included with the parameter set to 1.0 (i.e., an offset variable) or specified more generally as a covariate with a parameter to be estimated. The second option was used in this report. There is no reason to think that crashes will not increase linearly with segment length (an estimated parameter different than one is likely capturing the effect of one or more omitted variables that are correlated with segment length). Specifying the natural logarithm of segment length as a covariate and estimating its parameter may improve model prediction and reduce the standard error of the pavement marking parameter, which are desirable attributes given the objectives of this research. The disadvantage is that strict interpretation of a segment length parameter different than one may seem counterintuitive. As a sensitivity analysis, the log of segment length was also included as an offset variable and the analysis was repeated; the results did not change materially. (The coefficients for wider edge line as well as percent crash reduction estimates changed only slightly.)

Temporal correlations in the crash counts obtained from the same road segment over 6 years were handled by employing two different approaches: (1) negative binomial regression analysis on the crash frequencies aggregated over 6 years and (2) analysis on yearly crash frequencies using the negative binomial regression models with yearly trend and accounting for temporal correlations in the parameter estimation using the generalized estimating equations (GEEs) procedure. Similar conclusions were reached from both approaches. Only the results from the first approach (analyzing the aggregated crash counts over 6 years) are presented in this report.

Table 41 shows the estimates of the negative binomial regression model coefficients applied to Illinois non-intersection/interchange crashes during non-winter months for 6,531 segments (1,732.8 mi) aggregated for 6 years (2001 through 2006) and percent crash reduction estimates (animal collisions included). The regression coefficient for wider edge line was negative and statistically significant at = 0.05, which indicates a positive safety effect of wider edge lines (i.e., a smaller number of crashes is associated with wider edge lines) for the following crash types: F+I, daytime, daytime F+I, wet, single-vehicle F+I, and fixed object crashes. Lane departure crashes obtained as the sum of fixed object, head-on, and sideswipe crashes were also analyzed, and a positive safety effect of wider edge lines was observed as well.

Percent crash reduction estimates were computed by [1 - Exp(βedge)] x 100, where βedge represents the estimated coefficient of wider edge line. It can also be observed that the signs of the coefficients for lane width, shoulder width, log of AADT, and curve presence were consistent with intuition in most cases. For example, the negative signs of lane width and shoulder width coefficients imply that crashes tend to decrease as lane width or shoulder width increases, and the positive sign of curve presence implies that crashes tend to increase when there is a curve or curves as compared to when there is no curve.

Table 41. Estimates of regression coefficients of negative binomial regression models applied to Illinois
non-intersection/interchange crashes with animal collisions (2001-2006).


Crash Type

Intercept

Wider Edge Line

Lane Width

Shoulder Width

Log (AADT)

Presence of Curve

Log (Length)

Dispersion

Pearson Chi-Square/DF

Percent Crash Reduction

Total

-5.1248
(0.2797)

0.0077
(0.0442)

-0.0665
(0.0198)

-0.0183
(0.0062)

0.6878
(0.0257)

0.2038
(0.0550)

0.8197
(0.0148)

0.3253
(0.0236)

1.1248

-0.8

F+I

-6.9810
(0.5120)

-0.3792
(0.0885)

-0.0727
(0.0353)

-0.0499
(0.0109)

0.7846
(0.0456)

0.6199
(0.0916)

0.8614
(0.0267)

0.4003
(0.0658)

1.0813

31.6

PDO

-5.2852
(0.3158)

0.1127
(0.0488)

-0.0668
(0.0223)

-0.0089
(0.0069)

0.6637
(0.0290)

0.0432
(0.0655)

0.8083
(0.0167)

0.3814
(0.0299)

1.1197

-11.9

Day

-7.4326
(0.4046)

-0.2141
(0.0656)

-0.0956
(0.0277)

-0.0452
(0.0084)

0.9315
(0.0358)

0.1564
(0.0768)

0.7685
(0.0198)

0.3242
(0.0396)

1.0819

19.3

Night

-4.8206
(0.3683)

0.0931
(0.0577)

-0.0389
(0.0262)

0.0007
0.0082)

0.5044
(0.0342)

0.2737
(0.0773)

0.8746
(0.0205)

0.3756
(0.0399)

1.0935

-9.8

Daytime F+I

-7.6401
(0.6490)

-0.4580
(0.1161)

-0.1370
(0.0448)

-0.0612
(0.0136)

0.9040
(0.0574)

0.4398
(0.1208)

0.8381
(0.0329)

0.3948
(0.0955)

1.0994

36.7

Nighttime F+I*

-7.9660
(0.7893)

-0.2295
(0.1311)

0.0316
(0.0538)

-0.0359
(0.0164)

0.6161
(0.0681)

0.9747
(0.1361)

0.9293
(0.0411)

 

1.0554

20.5

Wet

-7.3318
(0.7675)

-0.2936
(0.1309)

-0.0684
(0.0537)

-0.0435
(0.0165)

0.7059
(0.0688)

0.3372
(0.1478)

0.8033
(0.0395)

0.9072
(0.1747)

1.0208

25.4

Wet nighttime*

-6.4462
(0.9922)

-0.3297
(0.1809)

-0.1006
(0.0707)

-0.0212
(0.0223)

0.5313
(0.0909)

0.4629
(0.2078)

0.8524
(0.0544)

 

1.0673

28.1

Single-vehicle

-3.9282
(0.3034)

0.0394
(0.0482)

-0.0406
(0.0218)

-0.0145
(0.0069)

0.4692
(0.0285)

0.3274
(0.0609)

0.8544
(0.0169)

0.3941
(0.0299)

1.1383

-3.9

Single-vehicle wet

-6.1985
(0.8318)

-0.2440
(0.1414)

-0.0324
(0.0593)

-0.0516
(0.0186)

0.4927
(0.0761)

0.4512
(0.1643)

0.8430
(0.0452)

1.1907
(0.2354)

1.0138

21.7

Single-vehicle nighttime

-4.6353
(0.3795)

0.1181
(0.0593)

-0.0239
(0.0271)

0.0031
(0.0086)

0.4483
(0.0355)

0.3303
(0.0795)

0.8873
(0.0215)

0.4057
(0.0436)

1.1028

-12.5

Single-vehicle F+I

-5.1562
(0.5837)

-0.4017
(0.1046)

-0.0258
(0.0413)

-0.0646
(0.0135)

0.4477
(0.0540)

0.8342
(0.1066)

0.8760
(0.0330)

0.5748
(0.1050)

1.0747

33.1

Single-vehicle nighttime F+I*

-7.3570
(0.8462)

-0.2171
(0.1411)

0.0516
(0.0582)

-0.0472
(0.0182)

0.4951
(0.0741)

1.1010
(0.1454)

0.9390
(0.0456)

 

1.0567

19.5

Older driver
(≥ 55 years old)

-7.2419
(0.5254)

-0.0168
(0.0818)

-0.0758
(0.0363)

-0.0213
(0.0110)

0.7691
(0.0465)

0.1166
(0.1092)

0.7871
(0.0264)

0.1733
(0.0595)

1.0639

1.7

Fixed object

-6.4892
(0.5880)

-0.3495
(0.1009)

-0.0307
(0.0408)

-0.0889
(0.0130)

0.6433
(0.0530)

0.4572
(0.1079)

0.7786
(0.0306)

0.6207
(0.1009)

1.1105

29.5

Lane departure

-7.1562
(0.5283)

-0.2429
(0.0875)

-0.0845
(0.0365)

-0.0674
(0.0114)

0.8225
(0.0474)

0.3958
(0.0967)

0.7855
(0.0268)

0.5112
(0.0757)

1.1230

21.6

*These crashes were fitted by Poisson regression models because negative binomial regression models could not be fitted due to insufficient data. Poisson regression models do not have dispersion parameters.
Note: Standard errors are provided in parenthesis. Results showing significant effects (at α = 0.05) are in bold.

The coefficients of the wider edge line for some of the crash types in table 41 (total, PDO, night, single-vehicle, and single-vehicle nighttime crashes) were positive, which indicates a negative safety effect of wider edge lines for those crash types. Researchers suspected that these counter-intuitive negative safety effects of wider lines on total, PDO, night, single-vehicle, and single-vehicle nighttime crashes were because of a high proportion of irrelevant crashes (animal collisions) in those types of crashes and reanalyzed the data after removing animal collisions.

The results of the analysis after the removal of animal collisions are presented in table 42. The table shows estimates of regression coefficients of negative binomial regression models applied to Illinois non-intersection/interchange crashes in non-winter months without animal collisions for 6,531 segments (1,732.8 mi) aggregated for 6 years (2001 through 2006) and percent crash reduction estimates. It can be observed from the table that the coefficients of the wider edge for all crash types were negative and statistically significant, which suggests a positive safety effect of wider lines. As expected, inclusion of irrelevant crashes, such as animal collisions, in the safety evaluation of wider lines can lead to erroneous results when the proportion of such irrelevant crashes in the data is non-negligible. The safety analysis results for F+I crashes were not significantly affected because only 10 percent of F+I crashes were animal collisions.

For Illinois, raised pavement markers (RPMs) are used statewide, as well as rumble strips on interstates. Discussions with Illinois Department of Transportation staff indicated that RPM and rumble strip use was not correlated with the presence of wider lines (i.e., the presence of RPMs and rumble strips was not more or less likely on roads with wider lines). It needs to be noted, however, that information on additional delineation and guidance measures (other than RPMs and rumble strips) was not available and could not be incorporated into the analysis. Therefore, the above observations are based on the same assumption made for rumble strips and RPMs that the effects of the variables not in the database such as additional delineation/guidance measures are the same (or averaged out) for the segments with and without wider edge lines. Finally, it should be restated that the pavement marking widths in Illinois vary by district. The pavement marking variable in the model may capture other differences between districts that also influence safety. Possible examples include differences in crash reporting (likely more significant at the minor to non-injury level), terrain differences, and roadside differences. This limitation is relevant to most cross-sectional studies.

The research team obtained Illinois crash data for two additional years, 2007 and 2008. While the 2001 through 2008 data covered a longer time period, the number of segments that were defined with exactly the same roadway geometric variable values throughout the observation period and the roadway mileage that they cover (643.3 mi) was considerably smaller compared to the 2001 through 2006 data (1,732.8 mi). As a result, the actual number of crashes contained in the 2001 through 2008 data was considerably smaller than the 2001 through 2006 data (see table 43). The research team determined that the 2001 through 2006 data gave the optimal coverage in terms of the amount of crash data when both time and space were considered.

Table 42. Estimates of regression coefficients of negative binomial regression models applied to Illinois
non-intersection/interchange crashes without animal collisions (2001-2006).


Crash Type

Intercept

Wider Edge Line

Lane Width

Shoulder Width

Log
(AADT)

Presence of Curve

Log (Length)

Dispersion

Pearson Chi-Square/DF

Percent Crash Reduction

Total

-7.9368
(0.3952)

-0.3587
(0.0659)

-0.0651
(0.0269)

-0.0579
(0.0083)

0.9801
(0.0350)

0.3460
(0.0690)

0.7714
(0.0192)

0.4334
(0.0402)

1.0907

30.1

F+I

-7.4089
(0.5477)

-0.4727
(0.0975)

-0.0861
(0.0377)

-0.0566
(0.0117)

0.8471
(0.0486)

0.6968
(0.0947)

0.8505
(0.0282)

0.4701
(0.0759)

1.0944

37.7

PDO

-9.6705
(0.5154)

-0.2728
(0.0823)

-0.0500
(0.0344)

-0.0599
(0.0103)

1.0996
(0.0443)

0.0703
(0.0933)

0.7241
(0.0235)

0.4688
(0.0590)

1.0821

23.9

Day

-9.0662
(0.4810)

-0.3438
(0.0791)

-0.0809
(0.0323)

-0.0595
(0.0097)

1.0878
(0.0417)

0.1721
(0.0861)

0.7487
(0.0224)

0.4135
(0.0520)

1.0893

29.1

Night

-8.0035
(0.6469)

-0.3559
(0.1103)

-0.0371
(0.0442)

-0.0614
(0.0137)

0.8108
(0.0571)

0.7443
(0.1104)

0.8567
(0.0332)

0.6157
(0.1069)

1.0611

29.9

Daytime F+I

-7.9157
(0.6718)

-0.4468
(0.1191)

-0.1350
(0.0461)

-0.0670
(0.0141)

0.9324
(0.0592)

0.4647
(0.1228)

0.8332
(0.0338)

0.4403
(0.1033)

1.1022

36.0

Nighttime F+I*

-8.8646
(0.8947)

-0.4186
(0.1581)

0.0185
(0.0606)

-0.0389
(0.0182)

0.7231
(0.0759)

1.1627
(0.1432)

0.9291
(0.0456)

 

1.0926

34.2

Wet

-9.4281
(0.9938)

-0.4260
(0.1706)

-0.0453
(0.0677)

-0.0884
(0.0207)

0.9129
(0.0865)

0.5833
(0.1666)

0.7836
(0.0482)

1.4959
(0.2962)

1.0229

34.7

Wet nighttime*

-8.6417
(1.4905)

-0.4419
(0.2689)

-0.0800
(0.1021)

-0.1039
(0.0321)

0.7494
(0.1301)

1.0403
(0.2471)

0.8765
(0.0780)

 

1.0695

35.7

Single-vehicle

-5.8930
(0.4684)

-0.4616
(0.0832)

-0.0241
(0.0328)

-0.0739
(0.0105)

0.6185
(0.0426)

0.6815
(0.0827)

0.8286
(0.0250)

0.5598
(0.0646)

1.1010

37.0

Single-vehicle wet

-7.7587
(1.0373)

-0.3968
(0.1804)

-0.0037
(0.0702)

-0.1279
(0.0226)

0.6380
(0.0902)

0.8635
(0.1790)

0.8695
(0.0540)

 

1.1688

32.8

Single-vehicle nighttime

-7.7545
(0.7084)

-0.3492
(0.1208)

-0.0055
(0.0486)

-0.0681
(0.0152)

0.6964
(0.0629)

0.9782
(0.1184)

0.8996
(0.0376)

0.6547
(0.1307)

1.0685

29.5

Single-vehicle F+I

-5.3920
(0.6376)

-0.5479
(0.1201)

-0.0394
(0.0451)

-0.0776
(0.0149)

0.4844
(0.0590)

0.9654
(0.1120)

0.8644
(0.0359)

0.7124
(0.1295)

1.0939

42.2

Single-vehicle nighttime F+I*

-8.2466
(0.9811)

-0.4504
(0.1764)

-0.0455
(0.0669)

-0.0524
(0.0206)

0.5873
(0.0845)

1.3439
(0.1553)

0.9432
(0.0518)

 

1.0879

36.3

Older driver
(≥ 55 years old)

-10.7785
(0.7606)

-0.2764
(0.1252)

-0.0699
(0.0507)

-0.0404
(0.0146)

1.1225
(0.0629)

0.1637
(0.1385)

0.7074
(0.0337)

 

1.0552

24.1

Fixed object

-6.4892
(0.5880)

-0.3495
(0.1009)

-0.0307
(0.0408)

-0.0889
(0.0130)

0.6433
(0.0530)

0.4572
(0.1079)

0.7786
(0.0306)

0.6207
(0.1009)

1.1105

29.5

Lane departure

-7.1562
(0.5283)

-0.2429
(0.0875)

-0.0845
(0.0365)

-0.0674
(0.0114)

0.8225
(0.0474)

0.3958
(0.0967)

0.7855
(0.0268)

0.5112
(0.0757)

1.1230

21.6

*These crashes were fitted by Poisson regression models because negative binomial regression models could not be fitted due to insufficient data. Poisson regression models do not have Dispersion parameters.
Note: Standard errors are provided in parenthesis. Results showing significant effects (at α = 0.05) are in bold.

Table 43. Summary of Illinois crash data for 2001 through 2008.


Variable

Dataset With Animal Collisions

Dataset Without Animal Collisions

4 inches

5 inches

4 inches

5 inches

4 inches

5 inches

4 inches

5 inches

Number of segments

2,572

642

2,572

642

2,572

642

2,572

642

Total segment length (mi)

520.0

123.3

520.0

123.3

520.0

123.3

520.0

123.3

Mile-years

4,160.0

986.4

4,160.0

986.4

4,160.0

986.4

4,160.0

986.4

Average AADT

3,160.2

2,248.4

3,160.2

2,248.4

3,160.2

2,248.4

3,160.2

2,248.4

Crash Type

Crash Counts

Crash Rates

Crash Counts

Crash Rates

Total

3,280

695

1.17

1.47

1,753

248

0.63

0.53

F+I

814

103

0.29

0.22

744

88

0.27

0.19

PDO

2,466

592

0.88

1.25

1,009

160

0.36

0.34

Daytime

1,472

247

0.53

0.52

1,134

171

0.41

0.36

Nighttime

1,534

362

0.55

0.77

541

64

0.19

0.14

Daytime F+I

488

55

0.17

0.12

468

52

0.17

0.11

Nighttime F+I

292

47

0.10

0.10

248

35

0.09

0.07

Wet

365

62

0.13

0.13

260

38

0.09

0.08

Wet nighttime

169

27

0.06

0.06

98

11

0.04

0.02

Single-vehicle

2,568

604

0.92

1.28

1,046

157

0.37

0.33

Single-vehicle wet

288

54

0.10

0.11

183

30

0.07

0.06

Single-vehicle nighttime

1,445

355

0.52

0.75

454

57

0.16

0.12

Single-vehicle F+I

551

77

0.20

0.16

483

62

0.17

0.13

Single-vehicle nighttime F+I

258

43

0.09

0.09

215

31

0.08

0.07

Older driver
(≥ 55 years old)

707

147

0.25

0.31

369

49

0.13

0.10

Fixed object

654

105

0.23

0.22

654

105

0.23

0.22

Note: Crash rates are computed as crashes per million vehicle miles of travel per year.

Although the focus of the study was rural two-lane roadways, the research team also compiled the Illinois freeway crash data from 2001 through 2006 from 571 segments (708 mi), of which 514 segments (593 mi) have a standard line width (4-inch edge lines and 4-inch skip lines), 13 segments (21 mi) have 4-inch edge lines and 6-inch skip lines, and 44 segments (94 mi) have 5-inch edge lines and 5-inch skip lines. Appendix D provides a summary of crash rates for those 571 freeway segments. Table 93 in appendix D shows that the freeway crash rates associated with wider skip lines or edge lines are, in general, lower than those associated with 4-inch edge lines and skip lines. However, for single-vehicle and single-vehicle nighttime crashes, slightly higher crash rates were observed for wider lines compared to 4-inch edge lines. Unfortunately, available sample sizes did not allow the development of negative binomial or Poisson regression models, which take confounders into account.

Analysis of Kansas Rural Two-Lane Roadway Crash Data

The Kansas crash data consist of non-intersection/interchange crash counts during non-winter months from 2,767 rural two-lane road segments (2,178.2 mi) in districts 2 and 6 from 2001 through 2008. An EB approach was employed to analyze the Kansas crash data. The EB method accounts for the effect on crash frequencies of regression to the mean along with changes in traffic volume and other changes in crash frequencies not due to the treatment. It has been considered a statistically defensible safety evaluation tool in observational before-after studies for more than two decades. In the EB method, safety performance functions (SPFs) are used to estimate the expected crash frequencies at the treated sites had treatments not been applied. Generalized linear regression models, specifically negative binomial regression models, are often used to derive the SPFs. The steps of the EB procedure used for the Kansas data analysis in this project are described below. Note that SPFs were calibrated for each year of the before and after periods rather than just for each period.

  1. Develop an SPF and estimate the regression coefficients and a negative binomial dispersion parameter (k) using data from the reference group.
  2. Estimate the expected number of crashes E(κiy) for each year in the before period at each treatment site using the SPF developed in step 1.
  3. Compute the sum of the annual SPF predictions during the before period at each treatment site using the equation in figure 16.

    . P subscript i equals the summation from y equals 1 to y subscript 0i minus 1 of E-hat open parenthesis kappa subscript iy close parenthesis.
    Figure 16. Equation. Predicted number of crashes in before period

    Where y0i denotes the year during which the countermeasure was installed at site i.

  4. Obtain an estimate of the expected number of crashes (Mi) before implementation of the countermeasure at each treatment site and an estimate of variance of Mi. The estimate Mi is given by combining the sum of the annual SPF predictions during the before period (Pi) with the total count of crashes (Ki) during the before period using the equation in figure 17.

    M subscript i equals w subscript i times P subscript i plus open parenthesis (1 minus w subscript i) close parenthesis times K subscript i.
    Figure 17. Equation. Expected number of crashes in before period

    Where Ki is the total crash count during the before period at site i and the weight wi is given by the equation in figure 18.

    w subscript i equals 1 divided by 1 plus k times P subscript i.
    Figure 18. Equation. Weight

    Where k is the estimated dispersion parameter of the negative binomial regression model developed in step 1. An estimated variance of Mi is given by the equation in figure 19.

    The variance of open parenthesis M subscript i close parenthesis equals open parenthesis 1 minus w subscript i close parenthesis times M subscript i.
    Figure 19. Equation. Estimated variance in expected number of crashes in before period

  5. Determine SPF predictions Ê (kiy) for each year in the after period at each treatment site and compute Ci, the ratio of the sum of the annual SPF predictions for the after period (Qi) and the sum of the annual SPF predictions for the before period (Pi) using figure 20.

    C subscript i equals summation from y equals y subscript 0i plus 1 to Y of E-hat open parenthesis kappa subscript iy close parenthesis divided by summation from y¬ equals 1 to y subscript 0i minus 1 of E-hat open parenthesis kappa subscript iy close parenthesis equals Q subscript i divided by P subscript i.
    Figure 20. Equation. Ratio of the sum of the annual SPF predictions for the after period

  6. Obtain the predicted crashes (The greek symbol Pi with a circumflex on top.) and its estimated variance during the after period that would have occurred without implementing the countermeasure. The predicted crashes (The greek symbol Pi with a circumflex on top.) are given by the equation in figure 21.

    Pi-hat subscript i equals C subscript i times M subscript i.
    Figure 21. Equation. Predicted number of crashes in after period with no countermeasure

    The estimated variance of The greek symbol Pi with a circumflex on top. is given by the equation in figure 22.

    The variance of open parenthesis pi-hat subscript i close parenthesis equals C subscript i squared times the variance of open parenthesis M subscript i close parenthesis equals C subscript i squared times open parenthesis 1 minus w subscript i close parenthesis times M subscript i.
    Figure 22. Equation. Estimated variance of predicted crashes in after period

  7. Compute the sum of the predicted crashes over all sites in a treatment group of interest and its estimated variance by using the equations in figure 23 and figure 24.

    Pi-hat equals summation from i equals 1 to I of pi-hat subscript i.
    Figure 23. Equation. Sum of predicted crashes for all sites in a treatment group

    The variance of open parenthesis pi-hat close parenthesis equals the summation from i equals 1 to I of the variance of open parenthesis pi-hat subscript i close parenthesis.
    Figure 24. Equation. Variance of total predicted crashes for all sites in a treatment group

    Where I is the total number of sites in a treatment group of interest.

  8. Compute the sum of the observed crashes over all sites in a treatment group of interest by using the equation in figure 25.

    L equals the summation from i equals 1 to I of L subscript i.
    Figure 25. Equation. Sum of observed crashes for all sites in a treatment group

    Where Li is the total crash counts during the after period at site i.

  9. The index of effectiveness of the countermeasure is estimated by the equation in figure 26.

    Theta-hat equals L divided by pi-hat times open parenthesis 1 plus var times open parenthesis pi-hat close parenthesis divided by pi-hat squared close parenthesis.
    Figure 26. Equation. Estimated index of effectiveness of a countermeasure

    The percent change in the number of crashes at site i is given by 100(1 -The greek symbol Theta with a circumflex on top.). If the index of effectiveness is less than 1, then the countermeasure has a positive effect on safety.

  10. Compute the estimated variance and standard error of the estimated index of effectiveness and the approximate 95 percent confidence interval for θ, the index of effectiveness. The estimated variance and standard error of the estimated index of effectiveness are given by the equations in figure 27 and figure 28.

    The variance of open parenthesis theta-hat close parenthesis equals theta-hat squared times open parenthesis 1 divided by L plus the variance of open parenthesis pi-hat close parenthesis divided by pi-hat squared close parenthesis, all divided by open parenthesis 1 plus the variance of open parenthesis pi-hat close parenthesis divided by pi-hat squared close parenthesis squared.
    Figure 27. Equation. Estimated variance in estimated index of effectiveness

    s.e. open parenthesis theta-hat close parenthesis equals the square root of the variance of open parenthesis theta-hat close parenthesis.
    Figure 28. Equation. Standard error of estimated index of effectiveness

The approximate 95 percent confidence interval for θ is given by adding and subtracting 1.96 s.e.(The greek symbol Theta with a circumflex on top) from The greek symbol Theta with a circumflex on top. If the confidence interval contains the value 1, then no statistically significant effect has been observed. This does not mean that a safety effect does not exist, so all indices that were estimated are presented in this report to provide a complete picture of safety effects. A confidence interval less than 1 (i.e., the upper limit of the interval was less than 1) implies that the countermeasure had a significant positive effect (i.e., a reduction in crashes) on safety. A confidence interval greater than 1 (i.e., the lower limit of the interval was greater than 1) implies that the countermeasure had a significant negative effect (i.e., an increase in crashes) on safety.

While the success of an EB approach largely depends on reliable estimation of SPFs, it is often hard to identify a reference group that is similar enough to the treatment group. Originally, the researchers considered sites untreated during the 8 years of the study period, 2001-2008, for Kansas. In Kansas, the wider lines were installed in 2005 through 2008. Table 44 summarizes the number of segments and the corresponding mileage for each implementation year. There were only 42.1 mi of roadways (90 segments) in the database that could be used as a reference group (without wider edge lines installed until the end of 2008). The limited length of comparable roadway made it difficult to develop reliable SPFs. Researchers decided to use the segments for wider edge lines that were installed in 2008 as additional sites for a reference group and restricted the study period to 7 years (2001-2007) instead. The number of segments and mileage for the resulting reference group are 261 and 158.1 mi, respectively. Because the segments implemented in 2007 (173 segments corresponding to 110.1 mi) do not have any after period data, 173 segments were excluded from the EB before-after evaluation, which left two treatment groups in the evaluation-group 1, implemented in 2005, consisted of 1,213 segments (803.8 mi) and group 2, implemented in 2006, consisted of 402 segments (361.5 mi). Also note that there are 718 segments in table 44 for which the implementation year is unknown. The only information researchers know about those segments is that wider lines were installed sometime after July 2005. Researchers conducted two sets of analysis for Kansas: one that excluded those 718 segments (analysis 1, which had 1,615 segments (1,165 mi) of rural two-lane roadways) and a second analysis that included them (analysis 2, which had 2,333 segments (1909.9 mi) of rural two-lane roadways). For analysis 2, the 718 included segments were placed in group 1 with an assumed implementation year of 2005 under the expectation that it would lead to more conservative safety effectiveness estimates (i.e., the effects of wider lines will be underestimated if the effects of wider lines are either null or positive and overestimated if the effects are negative, which is unlikely but possible).

Table 44. Number of segments and miles for each implementation year of wider edge lines in Kansas.


Implementation Year

Number of Segments

Miles

Unknown

718

744.7

2005

1,213

803.8

2006

402

361.5

2007

173

110.1

2008

171

116.0

Not implemented until 2008

90

42.1

Total

2,767

2,178.2

Types of crashes analyzed included the following:

The negative binomial regression models with indicator variables for district (2 and 6) and year (2001-2007) to control for general trends, shoulder width, log(AADT), and log(segment length) as independent variables were employed to develop SPFs. Although some other roadway characteristic variables, such as lane width and speed limit, were also available in the database, there was not much variation in those variables, so they were not included in the model. Due to an issue of the small sample size, the coefficients for SPFs could be directly estimated only for total, PDO, nighttime, single-vehicle, and fixed object crashes (recall that these crash types were all restricted to non-intersection/interchange crashes during non-winter months). SPFs for other crash types were obtained by applying a multiplier, af , (computed as the number of crashes of a specific type divided by the total number of crashes for the reference group) to the SPF for total crashes as in Bahar et al.(58) The estimated coefficients for SPFs for total, PDO, nighttime, single-vehicle, and fixed object crashes and the multipliers (af ) for the crash types considered in this study are presented in table 45 and table 46, respectively.

Table 45. Estimates of coefficients for SPFs developed based on a reference group
consisting of 263 segments (158.1 mi) on rural two-lane roadways in Kansas.

Variable

Total

PDO

Nighttime

Single-Vehicle

Fixed Object

District

2

-3.6538

-3.2213

-3.5047

-5.9793

-5.8683

6

-4.3942

-4.2987

-4.5591

-5.6635

-6.1639

Year

2001

-0.2680

-0.3680

-0.4723

-0.0026

0.2467

2002

-0.4696

-0.4754

-0.5300

-0.5760

-1.2433

2003

-0.3080

-0.3020

-0.2082

-0.2770

-0.5571

2004

-0.2427

-0.1521

-0.2027

-0.1425

-0.1167

2005

-0.4324

-0.5194

-0.3902

-0.2484

-0.5254

2006

-0.2561

-0.3073

-0.5010

0.0505

0.0216

2007

0.0000

0.0000

0.0000

0.0000

0.0000

Shoulder width

-0.0483

-0.0138

-0.0552

-0.1569

-0.0808

Log(AADT)

0.5417

0.4355

0.4378

0.7190

0.6017

Log(length)

0.9344

0.9387

0.8666

1.0178

1.0476

Dispersion

0.2777

0.2913

0.4745

0.9151

0.2666

Pearson chi-square/DF

1.0700

1.1218

1.0759

1.1422

1.0663

Table 46. Ratio of the number of crashes of a specific type and the total number
of crashes for the reference group.


Crash Type
αf

Total

1.000

F+I

0.226

PDO

0.774

Daytime

0.375

Nighttime

0.493

Daytime F+I

0.127

Nighttime F+I

0.071

Wet

0.064

Wet nighttime

0.033

Single-vehicle

0.358

Single-vehicle F+I

0.167

Single-vehicle nighttime

0.108

Single-vehicle nighttime F+I

0.052

Fixed object

0.182

Table 47 and table 48 include the results of two EB before-after evaluations (analyses 1 and 2) based on the Kansas crash data. It can be observed from the tables that almost all crash types resulted in statistically significant (95 percent confidence level) crash reduction estimates with the exception of nighttime, nighttime F+I, wet nighttime, and single-vehicle nighttime F+I crashes. As expected, the percent crash reduction estimates were, in general, slightly smaller for analysis 2; however, the overall pattern stayed the same. It needs to be noted that single-vehicle road departure crashes were especially relevant target crashes for assessing the safety effects of wider edge lines. The results in table 47 and table 48 support consistent safety effects of wider edge lines for single-vehicle and associated disaggregate crash types (e.g., single-vehicle nighttime and single-vehicle F+I).

A sensitivity analysis that uses the yearly coefficients (as well as the coefficients for other variables) from the total crash SPF for PDO, nighttime, single-vehicle, and fixed object crashes (and applies the corresponding  to the calibrated model) to predict for those crash types was conducted. The results are also presented in table 47 and table 48. It can be observed that the results from the sensitivity analysis for PDO, nighttime, single-vehicle, and fixed object crashes did not change materially from those obtained by using their own SPF model coefficients in table 45.

Table 47. Results of EB before-after evaluations based on the Kansas crash data for analysis 1.


Crash Type

Crashes-Before Period

Crashes-After Period

Estimated Index of Effectiveness (Standard Error)

95 Percent Confidence Interval
for θ

Percent Crash Reductionb

Observed
(K)

EB Estimate
(M)

Observed
(L)

EB Estimatea
(The greek symbol Pi with a circumflex on top)

Total

2,776

2,420.79

1,021

1,234.07

0.827 (0.028)

(0.772, 0.882)

17.3

F+I

474

481.37

156

242.01

0.644 (0.053)

(0.541, 0.748)

36.6

PDO

2,302

1,935.39

865

987.19

0.876 (0.032)

(0.813, 0.939)

12.4

PDOc

2,302

1,897.28

865

970.44

0.891 (0.033)

(0.827, 0.955)

10.9

Day

823

807.45

293

406.34

0.721 (0.044)

(0.635, 0.807)

27.9

Nighttime

1,610

1,258.77

589

614.95

0.958 (0.043)

(0.874, 1.041)

4.2

Nighttime c

1,610

1,212.16

589

619.53

0.950 (0.042)

(0.869, 1.032)

5.0

Daytime F+I

256

268.65

80

135.09

0.592 (0.067)

(0.460, 0.724)

40.8

Nighttime F+I

186

152.24

68

76.72

0.886 (0.109)

(0.673, 1.099)

11.4

Wet

178

137.49

54

69.45

0.777 (0.107)

(0.568, 0.987)

22.3

Wet nighttime

82

69.95

27

35.32

0.764 (0.148)

0.474, 1.054)

23.6

Single-vehicle

784

738.42

273

368.97

0.739 (0.048)

(0.644, 0.834)

26.1

Single-
vehicle c

784

770.66

273

387.84

0.704 (0.044)

(0.617, 0.790)

29.6

Single-vehicle F+I

325

350.25

113

176.63

0.640 (0.061)

(0.519, 0.760)

36.0

Single-vehicle nighttime

299

235.32

98

118.77

0.825 (0.085)

(0.659, 0.991)

17.5

Single-vehicle nighttime F+I

126

110.87

46

55.87

0.823 (0.122)

(0.583, 1.063)

17.7

Fixed object

382

368.42

160

195.28

0.819 (0.067)

(0.688, 0.950)

18.1

Fixed object c

382

385.29

160

194.55

0.822 (0.067)

(0.691, 0.953)

17.8

aEB estimate is the predicted number of crashes during the after period where wider lines had not been installed.
bPercent crash reduction = 100(1 - θ).
cIndicates the results from the sensitivity analysis using the coefficients from the total crash SPF for prediction.
Note: Bold indicates statistically significant percent crash reductions at 95 percent confidence level.

Table 48. Results of EB before-after evaluations based on the Kansas crash data for analysis 2.


Crash Type

Crashes-Before Period

Crashes-After Period

Estimated Index of Effectiveness (Standard Error)

95 Percent Confidence Interval
for θ

Percent Crash Reductionb

Observed
(K)

EB Estimate
(M)

Observed
(L)

EB Estimatea
(The greek symbol Pi with a circumflex on top)

Total

4,319

3,757.87

1,739

2,042.63

0.851 (0.022)

(0.807, 0.895)

14.9

F+I

820

757.23

311

408.68

0.761 (0.045)

(0.673, 0.848)

23.9

PDO

3,499

2,906.11

1,428

1,562.46

0.914 (0.026)

(0.862, 0.966)

8.6

PDOc

3,499

2,914.32

1,428

1,585.65

0.900 (0.026)

(0.849, 0.951)

10.0

Day

1,413

1,291.15

571

699.05

0.817 (0.036)

(0.746, 0.887)

18.3

Nighttime

2,426

1,834.85

959

938.24

1.022 (0.036)

(0.951, 1.093)

-2.2

Nighttimec

2,426

1,848.47

959

1,003.78

0.955 (0.033)

(0.890, 1.020)

4.5

Daytime F+I

450

420.57

176

226.93

0.775 (0.060)

(0.658, 0.892)

22.5

Nighttime F+I

315

235.66

121

127.08

0.952 (0.088)

(0.780, 1.124)

4.8

Wet

291

213.0

96

115.14

0.834 (0.086)

(0.665, 1.002)

16.6

Wet nighttime

135

107.80

44

58.19

0.756 (0.115)

(0.531, 0.981)

24.4

Single-vehicle

1,313

1,251.26

499

694.67

0.718 (0.035)

(0.649, 0.787)

28.2

Single-vehiclec

1,313

1,214.06

499

655.69

0.761 (0.036)

(0.691, 0.831)

23.9

Single-vehicle F+I

529

542.24

205

292.49

0.701 (0.050)

(0.602, 0.799)

29.9

Single-vehicle nighttime

486

363.77

170

196.32

0.866 (0.068)

(0.733, 0.999)

13.4

Single-vehicle nighttime F+I

209

170.67

74

91.97

0.804 (0.094)

(0.619, 0.989)

19.6

Fixed object

629

606.10

275

353.12

0.779 (0.049)

(0.683, 0.874)

22.1

Fixed objectc

629

600.14

275

324.37

0.848 (0.053)

(0.744, 0.951)

15.2

aEB estimate is the predicted number of crashes during the after period where wider lines had not been installed.
bPercent crash reduction = 100(1 - θ).
cIndicates the results from the sensitivity analysis using the coefficients from the total crash SPF for prediction.
Note: Bold indicates statistically significant percent crash reductions at 95 percent confidence level.

Analysis of Michigan Rural Two-Lane Roadway Crash Data

The Michigan crash data consist of non-intersection/interchange crash counts during non-winter months obtained from 253 rural two-lane road segments (851.5 mi) from 2001 through 2007. In Michigan, the change from 4- to 6-inch edge lines was made on almost all State-owned systems in 2004. Table 49 shows the annual aggregated crash counts from the 253 segments for crash types considered in this study. Because 2004 was the installation year of wider lines, crashes from that year were excluded from the subsequent safety analysis.

Table 49. Annual aggregated crash counts over 253 segments (851.5 mi) of rural two-lane roadways in Michigan for 2001-2007.


Crash Type

2001

2002

2003

2004

2005

2006

2007

Total

1,012

1,068

1,188

1,202

980

1,096

1,115

F+I

146

166

144

158

134

113

139

PDO

866

902

1,044

1,044

846

983

976

Daytime

396

441

444

505

406

415

446

Nighttime

462

468

562

504

450

521

522

Daytime F+I

83

103

86

117

88

65

92

Nighttime F+I

42

44

48

35

40

38

39

Wet

110

103

134

115

50

96

72

Wet nighttime

48

54

65

47

20

51

41

Single-vehicle

832

879

1,009

1,014

838

968

978

Single-vehicle wet

88

84

114

98

37

82

61

Single-vehicle night

437

443

524

480

432

502

505

Single-vehicle F+I

80

99

90

95

92

70

88

Single-vehicle nighttime F+I

36

33

36

28

33

28

32

Originally, researchers attempted to conduct an EB before-after analysis on the Michigan data. However, the widespread switch from 4- to 6-inch edge lines on almost all State-owned roads (i.e., all facility types) in 2004 left almost no sites within Michigan as an available reference group/comparison group in the before-after safety evaluation. Although the SPFs could be developed based on the before period data, the general time trends in crash frequencies from before to after periods not due to wider lines could not be easily estimated. The researchers attempted to use the Illinois F+I data to estimate the change in underlying trends. Michigan intersection crashes were also tested as an alternative, but the trends of these crashes in the before period were opposite to those on Michigan rural two-lane highways. The lack of an appropriate reference group within the same state remained one of the main limitations of the EB analysis of the Michigan data.

Researchers employed an alternative approach to perform a safety evaluation of Michigan rural two-lane roadway crash data. The new approach was an interrupted time series design. (See references 59-63.) An interrupted time series design is a quasi-experimental method used to determine the impact of an intervention. Campbell and Ross indicated, "In the Interrupted Time-Series, the ‘causal’ variable is examined as an event or change occurring at a single time, specified independently of inspection of the data." (p. 41)(59) In this instance, the causal variable (intervention) is the installation of wider lines that took place statewide in 2004. A generalized linear segmented regression analysis was used as a statistical method for analyzing the data from the interrupted time series design. Specifically, a negative binomial regression model that introduces time as a variable to control for overall trend and intervention (installation of wider lines) as a variable to estimate the effect of the wider lines was utilized. For time, the years prior to the installation of wider lines were coded as negative integers starting at -1 in descending order, and the years after the installation of wider lines were coded as positive integers starting at 1 in ascending order. For intervention, years corresponding to the after period were coded "1," and years in the before period were coded "0." An additional variable, time after intervention, which was coded "0" before the intervention and (time-t0) where t0 is the year of the intervention, can also be included in the model to estimate a possible change in the trend (not just in the level) in the expected number of crashes. At road segment i, the log of expected number of annual crashes in year t(µit) can be expressed as shown in figure 29.

Log mu subscript it equals beta subscript 0 plus beta subscript 1 times time subscript t plus beta subscript 2 times intervention subscript t plus beta subscript 3 times time after intervention subscript t plus beta subscript 4 times X subscript i,4t ellipsis plus beta subscript k times X subscript i,kt
Figure 29. Equation. Negative binomial regression model for interrupted time series

Where Xi,kt is the value of the kth predictor variable measured at road segment i in time t.

The underlying assumption for the above model is that the relationship between the log mean annual crash count and time is linear within each segment of time period (i.e., for the time period before the intervention and independently for the time period after the intervention). The intercept, β0, represents the baseline level of the log mean annual crash count, and β1 represents the baseline trend that corresponds to the change in the log mean annual crash count that occurs with each year before the intervention. The coefficients β2 and β3 represent the level change (i.e., the change in the intercept) in the log mean annual crash count immediately after the intervention and the change in the trend (i.e., the change in the slope) in the log mean annual crash count after the intervention, respectively. The key parameters of interest are β2 and β3, which can measure the effects of intervention, while β0 and β1 play the role of controlling for baseline level and trend.

In addition to time, intervention, and time after intervention, lane width, terrain, log(AADT), log(segment length), and log(number of rainy days) were included as predictors in the negative binomial regression model for Michigan crash data. GEEs were employed as an estimation method to account for correlation in crash counts obtained for multiple years from the same segment.

Table 50 contains the estimated coefficients for negative binomial regression models considered and the corresponding percent crash reduction estimates where the GEE approach was used as an estimation method. Originally, an additional variable, time after intervention, had been included in the negative binomial regression models to estimate a possible change in the trend (not just in the level) in the expected number of crashes. However, time after intervention was not statistically significant for any of the crash types considered in the study and was consequently dropped from the models to facilitate the interpretation of the results. It can be observed from table 50 that for total, PDO, nighttime, wet, wet nighttime, single-vehicle, single-vehicle wet, and single-vehicle nighttime crashes, statistically significant crash reductions at the 95 percent confidence level were found.

In addition to the crash types reported in table 50, opposite direction crashes and additional disaggregated F+I crashes such as wet F+I, wet nighttime F+I, and single-vehicle wet F+I were analyzed. However, due to insufficient data (there were very few crashes of those types), model coefficients could not be estimated, and reliable crash reduction estimates could not be obtained.

Table 50. Results of interrupted time series analysis applied to the Michigan crash data from 253 segments (851.5 mi) of rural
two-lane roadways with 3 years (2001-2003) of pre-intervention and 3 years (2005-2007) of post-intervention data.


Crash Type

Intercept

Time

Intervention

Lane Width (inches)

Terrain

Log
(AADT)

Log
(Length)

Log (No. of Rainy Days)

Percent Crash Reduction*

95 Percent Confidence Interval for Percent Crash Reduction

Total

-3.4846

0.0782

-0.3204

-0.0977

0.1721

0.5205

1.0980

0.0769

27.4

(15.4, 37.7)

F+I

-8.6073

0.0050

-0.1668

-0.0379

0.1945

0.8216

1.0277

0.0056

15.4

(-21.5, 41.0)

PDO

-3.3755

0.0953

-0.3633

-0.1140

0.1700

0.4981

1.1149

0.1026

30.5

(18.1, 40.9)

Day

-3.9724

0.0512

-0.2271

-0.1008

0.2735

0.5638

1.0141

-0.0968

20.3

(0.2, 36.4)

Night

-5.4095

0.0984

-0.3666

-0.0474

0.1008

0.5666

1.1596

0.1228

30.7

(15.0, 43.5)

Daytime F+I

-8.6123

0.0031

-0.0860

-0.0801

0.2077

0.9254

0.9238

-0.1079

8.2

(-43.3, 41.2)

Nighttime F+I

-10.7416

0.0348

-0.2564

0.0780

0.1336

0.6430

1.2491

0.2353

22.6

(-50.3, 60.2)

Wet

-11.2267

0.1715

-1.1140

-0.0626

0.2183

0.4813

0.9848

1.2745

67.2

(45.2, 80.3)

Wet nighttime

-11.8302

0.2715

-1.4633

-0.0321

0.1819

0.4009

1.0133

1.3551

76.9

(57.2, 87.5)

Single-vehicle

-2.9988

0.1004

-0.3566

-0.1117

0.1917

0.4313

1.1665

1.1046

30.0

(17.7, 40.5)

Single-vehicle wet

-9.9483

0.2313

-1.3394

-0.1147

0.2328

0.3202

1.0439

1.3670

73.8

(55.8, 84.5)

Single-vehicle nighttime

-5.2232

0.0987

-0.3476

-0.0519

0.1062

0.5438

1.1694

0.1174

29.4

(13.4, 42.4)

Single-vehicle F+I

-6.0126

0.0062

-0.1056

-0.0671

0.1209

0.5717

1.2009

-0.1233

10.0

(-40.8, 42.5)

Single-vehicle nighttime F+I

-8.1645

-0.0016

-0.1023

-0.0382

0.1039

0.5420

1.2847

0.0835

9.7

(-85.7, 56.1)

*Percent crash reduction estimates are obtained by {1 -Exp(βI)} x 100 where βI represents the estimated coefficient of the intervention variable.
Note: Statistically significant results at 95 percent confidence level are shown in bold.

The researchers obtained crash data for rural two-lane roadways in Michigan for two additional years-2008 and 2009. Because of the changes on some road segments after 2007, the number of segments of which roadway characteristics stayed the same for the entire study period (2001-2009) was reduced to 238 segments (787.8 mi). Table 51 shows the annual aggregated crash counts from the 238 segments for crash types considered in this study. Because 2004 was the installation year of wider lines, crashes from 2004 were excluded from the subsequent safety analysis.

Table 51. Annual aggregated crash counts over 238 segments (787.8 mi) of rural two-lane roadways in Michigan for 2001-2009.


Crash Type

2001

2002

2003

2004

2005

2006

2007

2008

2009

Total

943

981

1,106

1,119

905

1,010

1,034

1,006

1,007

F+I

127

149

127

146

125

99

121

124

115

PDO

816

832

979

973

780

911

913

882

892

Daytime

374

398

411

459

373

381

408

417

365

Nighttime

427

431

525

481

419

482

487

441

495

Daytime F+I

71

90

74

109

80

55

79

95

80

Nighttime F+I

38

41

43

33

39

35

35

25

30

Wet

101

100

129

105

46

89

67

67

97

Wet nighttime

43

52

63

44

19

47

39

31

50

Single-vehicle

778

811

946

953

775

894

915

890

891

Single-vehicle wet

81

81

109

90

34

76

56

55

83

Single-vehicle nighttime

403

409

493

458

401

463

472

428

479

Single-vehicle F+I

73

90

80

91

88

60

77

79

63

Researchers performed another interrupted time series analysis with 9 years of data as a sensitivity analysis. The number of rainy days could not be included in the models for the extended time period because the data for that variable were not available after 2007. Table 52 contains the results for the crash data obtained from 238 segments for 2001-2009 where a GEE approach was used as an estimation method. The results did not materially change from those in table 50, although the magnitude of crash reduction moderately decreased compared to the results based on 2001-2007 data (except for F+I, daytime F+I, and wet nighttime crashes). In addition to the crash types reported in table 52, opposite direction crashes and additional disaggregated F+I crashes such as single-vehicle nighttime F+I, wet F+I, wet nighttime F+I, and single-vehicle wet F+I were also analyzed. Due to the insufficient data, model coefficients could not be estimated, and reliable crash reduction estimates could not be obtained.

The research team also compiled the Michigan freeway crash data for 2001-2007 from 508 segments (1,067.4 mi). Appendix D provides the annual aggregated crash counts from those 508 freeway segments as well as the results of interrupted time series analysis on the freeway crash data. No consistent or statistically significant safety effects of wider lines were observed for the Michigan freeway crash data.

Table 52. Results of interrupted time series analysis applied to the Michigan crash data from 238 segments (787.8 mi)
of rural two-lane roadways with 3 years (2001-2003) of pre-intervention and 5 years (2005-2009) of post-intervention data.


Crash Type

Intercept

Time (year)

Intervention

Lane Width (inches)

Terrain

Log (AADT)

Log (Length)

Percent Crash Reduction*

95 Percent Confidence Interval for Percent Crash Reduction

Total

-3.0916

0.0451

-0.2151

-0.1302

0.1737

0.5542

1.1074

19.4

(10.1, 27.6)

F+I

-8.0168

0.0132

-0.1754

-0.0118

0.1088

0.7572

1.0270

16.1

(-11.1, 36.6)

PDO

-2.9525

0.0490

-0.2186

-0.1399

0.1806

0.5310

1.1199

19.6

(9.8, 28.4)

Day

-4.0554

0.0264

-0.1277

-0.1296

0.2397

0.5589

1.0106

12.0

(-3.5, 25.2)

Night

-4.7448

0.0489

-0.2081

-0.0734

0.1138

0.5801

1.1665

18.8

(6.3, 29.6)

Daytime F+I

-8.5857

0.0560

-0.2617

-0.0150

0.1331

0.7860

0.9419

23.0

(-6.6, 44.4)

Nighttime F+I

-9.1065

-0.0490

0.0560

0.0072

0.0614

0.6744

1.2300

-5.8

(-86.8, 39.5)

Wet

-5.2136

0.1185

-0.9847

-0.0808

0.1628

0.5140

0.9940

62.6

(45.6, 74.4)

Wet nighttime

-12.1894

0.2982

-1.5695

0.0253

0.1669

0.3807

1.0200

79.2

(60.2, 89.1)

Single-vehicle

-2.4692

0.0540

-0.2066

-0.1425

0.1984

0.4615

1.1610

18.7

(8.9, 27.4)

Single-vehicle wet

-3.4974

0.1386

-1.0768

-0.1574

0.1754

0.3824

1.0386

65.9

(48.6,77.4)

Single-vehicle nighttime

-4.5967

0.0511

-0.1983

-0.0807

0.1234

0.5624

1.1767

18.0

(5.2, 29.0)

Single-vehicle F+I

-5.8489

-0.0190

0.0191

-0.0339

0.0916

0.4476

1.1011

-1.9

(-44.8, 28.3)

*Percent crash reduction estimates are obtained by {1 -Exp(βI)} x 100 where βI represents the estimated coefficient of the intervention variable.
Note: Statistically significant results at 95 percent confidence level are shown in bold.

Consolidated Results

Table 53 presents consolidated results for estimations in the percent crash reductions from the six separate analyses. Note that while only the non-intersection/interchange non-winter crashes were considered for all three States, animal collisions were removed only from the Illinois data. Additionally, the animal collisions were excluded from the Kansas single-vehicle crash categories by default since the Kansas crash types were coded so that single-vehicle animal collisions were separated from other single-vehicle crash types in the raw data. Single-vehicle animal collisions were not removed from the crash categories in the Michigan dataset or from non-single-vehicle crash categories in the Kansas dataset. The overall effect of these crashes on the safety effectiveness estimates for Kansas and Michigan was minimal given the before-after observational study design for these two States (as opposed to the cross-sectional study design for Illinois where different numbers of animal collisions at different locations across the State could mask the effect of wider lines on other crash types). Overall, the results in table 53 support consistent safety effects of wider edge lines on the (relevant) crashes considered.

Table 53. Percent crash reduction estimates for wider edge lines on rural two-lane highways based on the crash data from three States.


Crash Type

Percent Crash Reduction

Illinois
(With Animal Collisions)

Illinois
(Without Animal Collisions)

Kansas
(Analysis 1)

Kansas
(Analysis 2)

Michigan
(Analysis 1)

Michigan
(Analysis 2)

Total

-0.8

30.1

17.5

15.0

27.4

19.4

F+I

31.6

37.7

36.5

24.4

15.4

16.1

PDO

-11.9

23.9

12.3

8.6

30.5

19.6

Daytime

19.3

29.1

28.6

18.6

20.3

12.0

Nighttime

-9.8

29.9

3.7

-2.4

30.7

18.8

Daytime F+I

36.7

36.0

41.5

22.7

8.2

23.0

Nighttime F+I

20.5

34.2

12.7

5.8

22.6

-5.8

Wet

25.4

34.7

22.9

17.2

67.2

62.6

Wet nighttime

28.1

35.7

24.3

24.9

76.9

79.2

Single-vehicle

-3.9

37.0

27.0

28.7

30.0

18.7

Single vehicle wet

21.7

32.8

 

 

73.8

65.9

Single-vehicle nighttime

-12.5

29.5

18.4

14.1

29.4

18.0

Single-vehicle F+I

33.1

42.2

36.8

30.5

10.0

-1.9

Single-vehicle nighttime F+I

19.5

36.3

18.7

20.3

9.7

 

Older driver

1.7

24.1

 

 

 

 

Fixed object

29.5

29.5

19.0

22.4

 

 

Note: Statistically significant results at 95 percent confidence level are shown in bold. Blank cells represent inadequate data available to perform statistical testing.

CRASH SEVERITY ANALYSIS OF SINGLE-VEHICLE CRASHES

The results of the crash frequency analysis provided detailed evidence to suggest that wider edge lines are effective in reducing crashes on rural two-lane highways, especially with regard to relevant target crashes such as single-vehicle crashes and related disaggregate crashes (e.g., single-vehicle nighttime and single-vehicle F+I). The safety effects of wider edge lines, measured in terms of crash frequencies, were consistently positive and statistically significant using data from the three States. Crash severity is also an important component of road safety. Crash severity was partially addressed by the frequency analysis, with crash reductions estimated by severity level (e.g., F+I and PDO). This was still a frequency analysis. Confounding factors that influenced both frequency and severity may influence severity-related conclusions. The crash reduction parameters were also estimated independently. Estimated reductions in F+I and PDO crashes may not necessarily sum up to equal the estimated reductions in total crashes.(64) Finally, the levels of severity were highly aggregated. Crashes resulting in a fatality or any level of injury (i.e., incapacitating, non-incapacitating, or possible injury) were grouped into one severity category. Disaggregating the injury levels may provide additional insights to crash severity effects.

This research focused on an alternative approach to explore the impacts of wider lines on crash severity. The analysis estimated the effects of wider lines on crash severity given that a crash has occurred. The effects of traffic, roadway, and vehicle occupant factors that also influence severity are incorporated into the analysis. The data used for estimation were from the same rural two-lane highway segments in Illinois and Michigan used for the frequency analysis. The Kansas pavement marking data were not available until the final stages of this research project. Efforts focused on preparing the Kansas data for the EB analysis and executing the analysis; severity modeling of Kansas data was not conducted. The concentration was single-vehicle crashes, which was the focus of the following possible outcomes discussed by the research team, which ultimately became the motivation for conducting this research task:

The severity effects of wider lines were empirically modeled in order to explore these potential outcomes. Published research exists on the application of discrete choice models to explore crash severity.(66) Their use in applied safety research is relatively limited. The methodologies in the Highway Safety Manual are frequency-based.(55) Methods to predict changes in crash frequency by severity are included in some chapters. Default distributions for crash severity are also used in the Highway Safety Manual algorithms. Since the use of discrete choice models has been relatively limited, the research team conducted a literature review on the application of such models to explore crash severity (see appendix I).

Modeling Approach

The logit model is the most widely used discrete choice model because the choice probabilities take a closed form and are readily interpretable. In the multinomial logit model, the probability that crash n will have severity i is given by the equation in figure 30.

P subscript n open parenthesis i close parenthesis equals exponential open parenthesis beta subscript i times chi subscript n close parenthesis divided by the summation of series I of exponential open parenthesis beta subscript I times chi subscript n close parenthesis.
Figure 30. Equation. Probability of a given crash having a specified severity using multinomial logit model

Where Xn is a set of variables that will determine the crash severity, and βi is a vector of parameters to be estimated. Utility functions defining the severity likelihoods are defined in figure 31.

S subscript in equals beta subscript i times chi subscript n plus epsilon subscript in.
Figure 31. Equation. Function for severity likelihood

Where Σin is a set of error terms that account for unobserved variables. The error terms for each choice should follow independent extreme value distributions (also called Gumbel or type I extreme value). The key assumption is that the errors are independent of each other. This independence means that the unobserved portion of utility for one severity alternative is unrelated to the unobserved portion of utility for another severity alternative. If the unobserved portion of utility is correlated over alternatives, then there are three options: (1) use a different model that allows for correlated errors, such as nested logit or mixed logit model, (2) respecify the representative utility so that the source of the correlation is captured explicitly and thus the remaining errors are independent, or (3) use the logit model under the current specification of representative utility, considering the model to be an approximation.

This independence of irrelevant alternatives (IIA) assumption is an important issue for the application of the multinomial logit model. If IIA holds, the ratio of probabilities for any two alternatives is entirely unaffected by the systematic utilities of any other alternatives. Tests of IIA were developed by McFadden et al.(67) Under IIA, the ratio of probabilities for any two alternatives is the same whether or not other alternatives are available. As a result, if IIA holds in reality, then the parameter estimates obtained on the subset of alternatives will not be significantly different from those obtained on the full set of alternatives. A test of the hypothesis that the parameters on the subset are the same as the parameters on the full set constitutes a test of IIA.(68) The null hypothesis of the test is that the coefficients of variables are equal for full set alternatives and subset alternatives (i.e., IIA holds). The test statistic has a chi-square distribution with the degrees of freedom equal to the number of coefficients estimated in the constrained (subset) model. If the null hypothesis is not rejected, then the IIA assumption holds, and the multinomial logit model is appropriate. The researcher should explore the three options stated above if the null hypothesis is rejected (i.e., the IIA assumption does not hold). As shown in the following sections, the IIA assumption was confirmed for the severity models in this report, indicating that the multinomial logit model was appropriate.

The likelihood ratio index is used to assess the goodness of fit of the logit model. The statistic measures how well the model, with its estimated parameters, performs compared to a model in which all the parameters except for the constant are zero (which is usually equivalent to having no model at all). The likelihood ratio index is defined in figure 32.

Rho equals 1 minus LL open parenthesis beta close parenthesis divided by LL open parenthesis 0 close parenthesis.
Figure 32. Equation. Likelihood ratio index

Where LL(β) is the value of the log-likelihood function at the estimated parameters and LL(0) is its value when all the parameters are set equal to zero.

Description of Data

The data used for model estimation were crashes occurring on the same rural two-lane highway segments as those used for the Illinois cross-sectional analysis and the Michigan interrupted time series analysis. The database consisted of all 2002-2006 Illinois single-vehicle crashes occurring on these segments, except those that were animal collisions, and all Michigan single-vehicle crashes from 2001-2003 (i.e., the before period) and 2005-2007 (i.e., the after period). The 2001 Illinois crashes were not used because there was a significant amount of missing occupant-related data, which is important information for the severity model specifications. The final datasets consisted of 4,061 rural two-lane highway single-vehicle crashes in Illinois and 2,483 rural two-lane highway single-vehicle crashes in Michigan. Many of the same variables were available for both datasets. Detailed information on vehicle occupants other than the driver was only available for Illinois. Table 54 and table 55 provide definitions of the variables used in the model specifications for Illinois and Michigan, respectively.

Table 54. Illinois variable definitions.


Variable

Description

Male occupants

The number of male occupants (not including the driver) in the vehicle

Female occupants

The number of female occupants (not including the driver) in the vehicle

Back restraint use

Indicator variable for back restraint use (1 = there is at least one occupant in the back seat that is not wearing a seatbelt)

Front restraint use

Indicator variable for front restraint use (1 = there is at least one occupant in the front seat, other than the driver, who is not wearing a seatbelt)

Max occupant age: back

Maximum age of occupants in the back seat (equals zero if no occupants)

Max occupant age: front

Maximum age of occupants in the front seat (equals zero if no occupants)

Driver age

Driver age

Alcohol use

Indicator variable for driver alcohol use (1 = alcohol use is suspected)

Driver sex

Indicator variable for driver sex (1 = male)

Driver restraint use

Indicator variable for driver restraint use (1 = not wearing seatbelt)

Fixed object collision

Indicator variable for collision type (1 = fixed object; 0 = rollover)

Wet road

Indicator variable for road condition (1 = road surface is wet)

Snow road

Indicator variable for road condition (1 = road surface has snow)

Rainy weather

Indicator variable for weather (1 = raining)

Snowy weather

Indicator variable for weather (1 = snowing)

Foggy weather

Indicator variable for weather (1 = foggy)

Road debris

Indicator variable for road condition (1 = road surface has debris)

Lane width

Lane width (ft)

Shoulder width

Shoulder width (ft)

Speed limit

Speed limit (mi/h)

5-inch line width

Indicator variable for pavement marking width (1 = 5 inches)

Sharp curve

Indicator variable for horizontal curve presence (1 = horizontal curve sharper than 2.5 degrees)

AADT

Average annual daily traffic (thousand vehicles per day)

Table 55. Michigan variable definitions.


Variable

Description

Driver age

Driver age

Driver restraint use

Indicator variable for driver restraint use (1 = not wearing seatbelt)

Driver sex

Indicator variable for driver sex (1 = male)

Alcohol use

Indicator variable for driver alcohol use (1 = alcohol use is suspected)

Fixed object collision

Indicator variable for collision type (1 = fixed object; 0 = rollover)

Foggy weather

Indicator variable for weather (1 = foggy)

Rainy weather

Indicator variable for weather (1 = raining)

Snowy weather

Indicator variable for weather (1 = snowing)

Other weather condition

Indicator variable for weather (1 = other/unknown)

Daylight

Indicator variable for light condition (1 = daylight)

Wet road

Indicator variable for road condition (1 = road surface is wet)

Icy road

Indicator variable for road condition (1 = road surface has ice)

Snow road

Indicator variable for road condition (1 = road surface has snow)

Other road condition

Indicator variable for road condition (1 = other/unknown)

Speed limit

Speed limit (mi/h)

Shoulder width

Shoulder width (ft)

Lane width

Lane width (ft)

6-inch edge line width

Indicator variable for edge line width (1 = 6 inches)

AADT

Average annual daily traffic (thousand vehicles per day)

Results and Conclusions

General model specifications were used since this was an exploratory analysis of crash severity. Table 56 and table 57 provide the model estimation results using the Illinois and Michigan data, respectively. The PDO crash was set as the base outcome for both models. Generally, positive parameters indicate that the respective level of severity became more likely as the value for the variable increased. For example, the parameters for the driver restraint use indicator variables were positive, indicating that higher levels of severities become more likely if the driver was not wearing a seat belt. As expected, this variable is highly associated with crash severity in both the Michigan and Illinois models. Parameters for ADT were negative, indicating that higher levels of severities were less likely as traffic increased, likely due to slower travel speeds. Crash severity increased with driver age. The increased likelihood of injuries to older drivers likely offset the lower impact speeds that are more likely for those drivers. This is consistent with findings that sometimes show a U-shaped curve for driver fatalities per vehicle miles of travel by age, with an increasing trend from age 50 and above. This trend reflects an increased likelihood that crash involvement proves fatal more than an increase in crash frequencies. The signs are mixed for some variables. For example, the probability of a possible injury crash is lower, and the probabilities of non-incapacitating, incapacitating, and fatal crashes are higher as the number of male occupants in the vehicle increases. The findings for posted speed limit are also mixed. Increased severity is associated with higher posted speeds for all severities except fatalities in the Illinois model. The parameter estimate signs are not as consistent in the Michigan model. None of the parameters for posted speed limit are statistically significant in either model, likely reflecting that the impact speed of the crash as opposed to the posted speed is the most important speed-related variable.

Table 56. Multinomial logit model estimation results for severity of single-vehicle crashes on rural two-lane highways
in Illinois.

 

Variable

Possible Injury

Non-Incapacitating

Incapacitating

Fatal

Coefficients

p-value

Coefficients

p-value

Coefficients

p-value

Coefficients

p-value

Constant

-2.462

0.06

-0.279

0.716

-1.296

0.205

-2.565

0.219

Male occupants

-0.394

0.113

0.049

0.66

0.086

0.568

0.313

0.25

Female occupants

0.107

0.527

0.160

0.097

0.151

0.266

0.348

0.158

Back restraint use

1.185

0.194

-0.003

0.995

1.196

0.043

1.037

0.235

Front restraint use

1.329

0.073

1.34

0.002

1.459

0.002

2.409

0

Max occupant age: back

-0.024

0.532

0.016

0.212

-0.006

0.756

-0.073

0.23

Max occupant age: front

0.005

0.540

0.006

0.173

0.012

0.037

0.014

0.131

Driver age

0.001

0.789

0.002

0.419

0.016

0

0.034

0

Alcohol use

-0.178

0.565

0.816

0

0.883

0

1.46

0

Driver sex

-0.391

0.017

-0.329

0.001

-0.244

0.065

0.033

0.911

Driver restraint use

0.262

0.564

1.526

0

2.574

0

2.708

0

Fixed object collision

-0.625

0

-0.5

0

-0.519

0

0.252

0.36

Wet road

0.573

0.081

-0.091

0.669

0.209

0.431

-0.598

0.34

Snow road

0.175

0.525

-0.368

0.054

-0.579

0.058

-1.696

0.114

Rainy weather

-0.228

0.533

-0.036

0.877

-0.483

0.123

-0.319

0.67

Snowy weather

-0.448

0.087

-0.714

0

-1.091

0

-1.017

0.121

Foggy weather

-1.756

0.088

-0.187

0.554

-0.079

0.84

1.054

0.121

Road debris

-0.578

0.181

-0.626

0.013

-0.46

0.162

-0.549

0.47

Lane width

0.074

0.481

-0.008

0.902

-0.025

0.758

-0.211

0.214

Shoulder width

0.025

0.46

-0.037

0.065

-0.062

0.022

-0.02

0.723

Speed limit

0.174

0.554

0.207

0.207

0.144

0.497

-0.3

0.444

5-inch line width

-0.043

0.879

0.205

0.19

-0.099

0.659

-1.048

0.095

Sharp curve

0.289

0.225

0.174

0.235

-0.02

0.919

0.422

0.24

AADT

-0.141

0.001

-0.084

0

-0.061

0.026

-0.01

0.835

Note: Significant (α = 0.10) effects are shown in bold.

Table 57. Multinomial logit model estimation results for severity of single-vehicle crashes on rural, two-lane highways
in Michigan.


Variables

Possible Injury

Non-Incapacitating

Incapacitating

Fatal

 

Coefficients

p-Value

Coefficients

p-Value

Coefficients

p-Value

Coefficients

p-Value

Constant

-1.236

0.359

-0.321

0.824

-0.572

0.722

-60.637

0.993

Driver age

0.002

0.73

0.002

0.772

0.015

0.069

0.04

0.044

Driver restraint use

0.747

0.179

1.403

0.003

3.267

0

4.352

0

Driver sex

-0.329

0.065

-0.261

0.233

-0.419

0.173

-0.544

0.46

Alcohol use

-0.483

0.217

1.071

0

0.035

0.94

1.506

0.096

Fixed object collision

-0.666

0.001

-0.767

0.001

-0.99

0.002

-1.155

0.145

Foggy weather

-0.245

0.316

-0.119

0.665

-0.193

0.607

0.276

0.758

Rainy weather

-0.424

0.333

-0.144

0.792

0.717

0.293

2.836

0.148

Snowy weather

-0.072

0.83

-0.16

0.716

0.393

0.549

2.894

0.076

Other weather condition

0.283

0.525

-0.165

0.813

-13.85

0.989

-10.663

0.995

Daylight

-0.023

0.903

-0.149

0.511

-0.209

0.512

0.2

0.809

Wet road

-0.238

0.496

-0.516

0.223

-0.789

0.173

-1.622

0.302

Icy road

-0.765

0.01

-1.277

0.002

-1.631

0.007

-16.552

0.98

Snow road

-1.366

0.001

-1.401

0.009

-2.793

0.003

-3.275

0.079

Other road condition

-0.519

0.197

-0.348

0.48

-1.807

0.053

-16.365

0.988

Speed limit

0.008

0.967

-0.055

0.795

-0.271

0.201

9.654

0.994

Shoulder width

0.086

0.266

0.042

0.644

0.094

0.457

-0.213

0.436

Lane width

0.046

0.804

0.033

0.881

0.124

0.695

0.285

0.706

6-inch edge line width

-0.016

0.934

-0.495

0.024

-0.223

0.49

-0.619

0.421

AADT

-0.004

0.914

-0.049

0.256

-0.07

0.236

-0.347

0.11

Note: Significant (α = 0.10) effects are shown in bold.

The IIA assumption for all severity levels held true for both the Illinois and Michigan models, indicating that the multinomial logit model was appropriate. The signs of the parameters for edge line width were negative (except for non-incapacitating injury crashes in Illinois), indicating that crashes were less severe on road segments with wider lines. The only edge line parameters that were statistically significant (α = 0.1) were the parameters for fatal crashes in Illinois (indicating that the probability of a fatality, given that a crash occurred, was lower with wider lines) and for non-incapacitating injury crashes in Michigan (indicating that the probability of a non-incapacitating injury, given that a crash occurred, was lower with wider lines). Overall, the level of confidence was not high enough to make conclusive remarks on the effect of edge line width on crash severity. The patterns and signs of the marking width parameters do indicate either a reduction in severity or no severity effect, which supports an overall safety benefit of wider lines given the results of the frequency analysis. Table 58 and table 59 provide the severity distributions with and without wider markings in Illinois and Michigan, respectively. These distributions are predicted using the logit model results, so they are more powerful than a simple univariate comparison.

Table 58. Single-vehicle crash severity distributions with and without wider pavement markings on rural two-lane highways in Illinois.


Crash Type

4-Inch Lines
(percent)

5-Inch Lines
(percent)

Fatal

1.25

0.42

Incapacitating injury

10.34

9.08

Non-incapacitating injury

22.32

26.58

Possible injury

5.23

4.86

PDO

60.87

59.06

Table 59. Single-vehicle crash severity distributions with and without wider edge lines
on rural two-lane highways in Michigan.


Crash Type

4-Inch Lines (percent)

6-Inch Lines (percent)

Fatal

0.00

0.00

Incapacitating injury

2.14

1.81

Non-incapacitating injury

12.03

7.75

Possible injury

13.38

13.91

PDO

72.44

76.53

SUMMARY

Prior to this research, the results of work on the safety benefits of wider edge lines were inconclusive. The research reported herein provided a unique opportunity to explore the safety benefits of wider edge lines in the most comprehensive study on the topic to date.

Consolidated results for estimations in the percent crash reductions (six total analyses) support consistent safety effects of wider edge lines on the non-intersection/interchange non-winter crashes considered. Crash frequency analysis suggests that wider edge lines are effective in reducing crashes on rural two-lane highways, especially with regard to relevant target crashes such as single-vehicle crashes and related disaggregate crashes.

Generally, positive parameters indicated that the respective level of crash severity became more likely as the value for the variable increased, but results were mixed based on parameter. Statistically conclusive remarks on the effect of edge line width on crash severity cannot be made, but the patterns and signs of the marking width parameters support an overall safety benefit of wider lines given the results of the frequency analysis.

 

Federal Highway Administration | 1200 New Jersey Avenue, SE | Washington, DC 20590 | 202-366-4000
Turner-Fairbank Highway Research Center | 6300 Georgetown Pike | McLean, VA | 22101