Guidance for the Design and Application of Shoulder and Centerline Rumble Strips (2009)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

42 S E C T I O N 6 The safety evaluation of shoulder rumble strips addresses two key unresolved issues: (a) the safety effectiveness of shoul- der rumble strips on various roadway types, and (b) whether the placement of rumble strips with respect to the edgeline impacts the safety effectiveness of the treatment. Other issues that are addressed in lesser detail include determining the impact of shoulder rumble strips on heavy vehicle crashes, crashes that occur under adverse pavement conditions, and crashes that occur during low-lighting conditions. Previous safety evaluations of shoulder rumble strips have focused on determining the safety effectiveness of this treat- ment installed along freeway facilities, primarily in rural areas, but, to a lesser degree, in urban areas as well. This is mainly because the initial use of shoulder rumble strips was prima- rily on rural freeways. As evident from the survey results in Section 5, in recent years shoulder rumble strips have been installed along all types of roadways. NCHRP Report 617 (28) indicates that rolled shoulder rumble strips reduce all SVROR crashes by 21 percent on rural freeways and by 18 percent on all freeways (i.e., both rural and urban combined). It also reports that rolled shoulder rumble strips reduce SVROR- injury crashes by 7 percent on rural freeways and by 13 per- cent on all freeways. These safety estimates and others, including safety estimates for rural multilane divided high- ways, are incorporated in the draft chapters of the forthcom- ing HSM, but in most cases the safety estimates (i.e., AMFs) are not very reliable. In summary, reliable estimates of the safety effectiveness of shoulder rumble strips for different roadway types are not available, and in all likelihood the safety benefits of shoulder rumble strips vary by roadway type because the different types of roadways are built to varying standards (e.g., lane widths, shoulder widths, roadside), accommodate varying traffic volumes and distributions, serve different driver populations, and accommodate a range of operating speeds. Also evident from the survey results, transportation agen- cies install rumble strips at various locations with respect to the edgeline. Typical offset distances range from 0 to 30 in. (0 to 762 mm) from the edgeline. By placing the shoulder rumble strips closer to the edgeline (or in some cases on the edgeline), drivers are alerted sooner that their vehicles have departed from the travel lane than if the rumble strips are placed further from the edgeline on the shoulder. More recovery area is also available on the shoulder when the rum- ble strips are located closer to the edgeline. On the other hand, when rumble strips are located closer to the edgeline, drivers are more likely to run over them in nonemergency situations. It is not known how the offset distance influences the safety effectiveness of rumble strips. Other issues that have not been fully investigated in previ- ous research concern the impact that shoulder rumble strips may have on specific target crashes, such as heavy vehicle crashes and/or crashes that occur under adverse pavement conditions or low-lighting (i.e., nighttime) conditions. The safety effect that shoulder rumble strips have on crashes involving heavy vehicles is of interest because (a) it is unclear whether the stimuli (i.e., noise and vibration) generated by rumble strips are sufficient to alert drivers of heavy vehicles, (b) designing rumble strip patterns specifically for heavy vehicles will likely conflict with needs of other road users such as bicyclists, and (c) it is difficult to assess the need or prior- ity to specifically consider heavy vehicles in the design of shoulder rumble strips given the frequency of crashes involv- ing heavy vehicles that would likely be affected by the instal- lation of shoulder rumble strips. Finally, although the primary purpose of shoulder rumble strips is to alert inattentive and drowsy drivers that their vehicles have departed from the travel lanes, it is likely that this safety treatment also indirectly affects crashes that occur under adverse pavement conditions (i.e., rainy or snowy conditions) and during low-lighting (i.e., nighttime) conditions. For example, snow plow drivers have noted that they have come to depend on shoulder rumble strips to help them find the edge of the travel lane during heavy snow and other low-visibility situations, so in some sit- Safety Effectiveness of Shoulder Rumble Strips

uations when the lane lines (i.e., edgelines) might be difficult to see either because of precipitation or low retroreflectivity, shoulder rumble strips can serve to provide positive guid- ance to drivers who are not necessarily inattentive or drowsy, but may just find it difficult to follow the delineation of the roadway. This section describes the general scope of the safety eval- uation conducted to resolve these issues, the site selection process, the videolog data collection procedures, the database development, the analysis approach, and the analysis results. Scope of Safety Evaluation The primary objectives of the safety evaluation conducted as part of this research are to do the following: • Quantify the safety effectiveness of milled shoulder rumble strips on specific types of roads including urban freeways, urban multilane divided highways (nonfreeways), urban multilane undivided highways (nonfreeways), urban two- lane roads, rural freeways, rural multilane divided high- ways (nonfreeways), rural multilane undivided highways (nonfreeways), and rural two-lane roads. • Quantify the safety effectiveness of shoulder rumble strips placed in varying locations with respect to the edgeline. The safety effectiveness evaluation is based on the change in crash frequency for total (TOT) crashes, fatal and injury (FI) crashes, SVROR crashes, and/or SVROR FI crashes. Depend- ing upon the comparison, the data are analyzed using either an Empirical Bayes (EB) methodology for before-after analy- sis or a cross-sectional approach for analyses across treatment sites (i.e., sites with milled shoulder rumble strips) and non- treatment sites (i.e., sites without any type of shoulder rumble strip). The Analysis Approach part of this section describes the similarities and differences between these two analysis approaches. Additional analyses are performed to investigate the impact that shoulder rumble strips have on selected tar- get SVROR crashes, including the following: • Crashes involving heavy vehicles (i.e., trucks), • Crashes occurring under adverse pavement conditions, and • Crashes occurring during low-light conditions (i.e., dusk, dawn, or dark). As part of the safety evaluation of shoulder rumble strips, steps were taken in an effort to quantify the difference in safety effectiveness between rumble strips installed on the right (outside) shoulder of a divided highway and rumble strips installed on the left (median) shoulder of a divided highway. A key element for being able to address this issue is the capability to distinguish the location of the crashes within the right-of-way. Rumble strips installed on the right (out- side) shoulder are expected to reduce SVROR crashes to the right that occur on the roadside, while rumble strips installed on the left (median) shoulder are expected to reduce SVROR crashes to the left that occur within the median. Rules were developed to distinguish between SVROR-right and SVROR- left crashes from the electronic crash databases that had been assembled in conjunction with this safety evaluation. The accuracy of the rules was assessed by comparing the query results to a sampling of hard copies of the crash reports. Approximately 100 crash reports were reviewed during this process. Based upon the results of the sampling, SVROR-left and SVROR-right crashes could not be accurately distin- guished in the electronic databases assembled. Because resources were too limited to review hard copies of the crash reports for all of the crashes (or a large sampling of the crashes) in the respective databases, a decision was made to terminate efforts to address this issue. Based upon the safety evaluation of milled shoulder rum- ble strips on a range of roadway types, AMFs are developed for potential incorporation in HSM. The AMFs (and associ- ated standard errors) are developed in a manner consistent with the method correction factor procedures developed for use in conjunction with the HSM (65). Site Selection Representatives from the Georgia, Kentucky, Minnesota, Missouri, and Pennsylvania DOTs were contacted to inquire about potentially including sites from their respective states in the safety evaluation. Types of information gathered to determine whether sites from a particular state would be appropriate for this study included the following: • Has the agency installed shoulder rumble strips on a range of roadway types of interest for this study? • Does the agency have the ability to identify locations where shoulder rumble strips have been installed? • Does the agency have the ability to provide construction history information such as rumble strip installation dates and information about other improvements made during the study period? • Does the agency keep a library of videologs that could be accessed by the research team? • Is the agency willing to participate in the research and work with the research team to gather the necessary data for the safety evaluation? Through a series of phone interviews and, in one case, a visit to the central office, it was determined to include sites from Minnesota, Missouri, and Pennsylvania in the safety evalua- 43

tion. Subsequently, official requests were made to the respec- tive DOTs to provide a list of locations where shoulder rum- ble strips had been installed. A list of selection criteria was provided and explained to the DOTs in an effort to develop a list of treatment sites that could be used in a before-after safety evaluation. The following were selection criteria for identify- ing candidate treatment sites: • Roadway type—Identify locations along the following roadway types where shoulder rumble strips have been installed: – Urban freeways, – Urban multilane divided highways (nonfreeways), – Urban multilane undivided highways (nonfreeways), – Urban two-lane roads, – Rural freeways, – Rural multilane divided highways (nonfreeways), – Rural multilane undivided highways (nonfreeways), and – Rural two-lane roads. • Installation date—Shoulder rumble strips should have been installed sometime between calendar years 1997 and 2003. • Type of safety improvement: – 1st priority—Sites where installation of a shoulder rum- ble strip was the only recent improvement (safety or otherwise) made to the site. – 2nd priority—Sites where the shoulders (and/or travel lanes) were paved in conjunction with installing shoul- der rumble strips (i.e., a resurfacing project followed by the installation of shoulder rumble strips). The shoul- der widths before and after paving should be the same, and no other improvements (safety or otherwise) have been made recently. • Type of rumble strip: – The focus of the study is on milled rumble strips, so a priority is placed on sites with milled rumble strips. • Placement of the shoulder rumble strips with respect to the edgeline: – Sites should be identified with a range of offset distances (i.e., from edgeline rumble strips/stripes to offsets as far from the edgeline as the DOT’s policy permits). Ideally, sites with a range of offsets within a roadway type (e.g., rural freeways) would be preferable. The information provided by each state DOT based upon this initial inquiry for a list of treatment sites varied consid- erably. The following sections summarize the tasks con- ducted for the respective states to select treatment sites (i.e., sites with milled shoulder rumble strips) for inclusion in this safety evaluation. The next part on videolog data collection provides more details on how treatment and nontreatment sites (i.e., sites without any type of shoulder rumble strip, whether milled, rolled, formed, or raised) were identified for use in the analysis. Minnesota Sites Minnesota DOT (MnDOT) provided an initial list of treat- ment locations. The information provided with this list included the district, route type, route number, beginning and ending mileposts, installation dates, notes concerning the type of rumble strip installation, and dimension and offset information. The initial list contained many duplicate sites, and much of the information was incomplete. A series of tele- phone interviews were held with each district office to do the following: • Learn more about the construction history of each site to determine whether the rumble strips were installed in con- junction with other improvements at the site or were a stand-alone improvement. In many cases, the shoulder rumble strips were installed as part of a resurfacing project, but the cross section of the roadway (e.g., lane widths and shoulder widths) remained unchanged. • Verify the type of rumble strip installation (e.g., milled shoulder rumble strips). In some cases, candidate sites from the list were eliminated from further consideration because they included rolled rumble strips rather than milled rumble strips. • Gather missing information, such as the installation dates of the rumble strips, and when possible, offset information. Following the interviews, a comprehensive list of treatment sites was compiled, eliminating duplicate sites and only includ- ing those sites that appeared to be the most appropriate for inclusion in a before-after evaluation of milled shoulder rum- ble strips. Using roadway characteristic data available from the FHWA’s Highway Safety Information System (HSIS), the treatment sites were categorized according to the eight road- way types of interest for the study. After having compiled a prioritized list of sites for data collection, the research team reviewed each of the sites using MnDOT’s videolog system. The next part on Videolog Data Collection provides detailed information on the actual data collection process performed for sites in Minnesota and the other states. MnDOT’s videolog system contains videologs for calendar years 2000–2006. Initially, the 2006 videologs were reviewed to confirm the presence or absence of the rumble strips at the locations. Subsequently, the 2001–2005 videologs were reviewed to confirm the installation dates of the rumble strips at each site where the rumble strips were installed during cal- endar years 2002–2005. In the prioritized list of candidate treatment sites, the installation dates of the rumble strips 44

45 ranged from 1983 to 2006. Only those installation dates between calendar years 2002 and 2005 could be confirmed during this process by verifying the absence/presence of the shoulder rumble strips across multiple years. This process of confirming installation dates revealed that approximately 50 percent of the dates obtained either from the initial list or during the interviews with district personnel were correct. This relatively poor accuracy level of the instal- lation dates for the treatment sites caused serious concern, especially since it is reasonable to assume that the more recent installations would have more accurate information than the older installations. As a result of this finding, the following rules were applied concerning installation dates and the applicability of a site for inclusion in a before-after evaluation or a cross-sectional analysis: • For those sites with installation dates prior to and includ- ing calendar year 2000, the site was automatically deter- mined to be inappropriate for a before-after evaluation but appropriate for a cross-sectional analysis, and the analysis period was defined to extend from 2001 to 2005. • For those sites with installation dates during calendar year 2001, if the rumble strips were confirmed to be present during the review of the 2001 videolog, then the site was determined to be inappropriate for a before-after evalua- tion but appropriate for a cross-sectional analysis, and the analysis period was defined to extend from 2001 to 2005. If the rumble strips were confirmed to be absent during the review of the 2001 videolog but were present during the review of the 2002 videolog, then the site was determined to be appropriate for a before-after evaluation, and the analysis period was defined to extend from 1997 to 2005 with an installation date of 2001. • For those sites with installation dates during calendar years 2002 and 2005, installation dates were confirmed or mod- ified based upon the absence and/or presence of the rum- ble strips from the yearly videologs. In some cases the videologs were recorded in the spring, while others were recorded in the fall so the date of the yearly videologs was taken into consideration when determining the calendar year of the installation. Through this process, the treatment sites in Minnesota were selected for inclusion in this safety evaluation, the sites were classified as being appropriate for a before-after evalua- tion and/or a cross-sectional analysis, and the analysis peri- ods were determined based upon the available crash data and construction history. More details are provided in Analysis Approach later in this section. Table 10 shows the total mileage (by roadway type) of treat- ment and nontreatment sites from Minnesota considered for inclusion in the safety evaluation. This table reflects total mileage for treatment and nontreatment sites after a series of data quality checks were performed to ensure data were con- sistent and complete for each location. This table does not classify the mileage by installation dates, or by total mile-years that can be used for before- and after-period analyses in a before-after evaluation or the total mile-years that can be used in a cross-sectional analysis. This level of detail is provided in the descriptive statistics part of this section. Missouri Sites Missouri DOT (MoDOT) did not have an efficient manner to identify locations where milled shoulder rumble strips were present based upon the selection criteria provided. MoDOT’s central office initially generated a list of locations where shoul- der rumble strips were installed during 2001 through 2005; however, this list did not provide information on whether the rumble strips were installed as the only improvement to the site or in conjunction with other improvements, nor did the list identify the type of rumble strip (i.e., milled, rolled, or formed). Each district office was contacted to help provide this information. The district offices provided the requested infor- Roadway type Treatment sites (mi) Nontreatment sites (mi) 79.01127.1syaweerfnabrU Urban multilane divided highways (nonfreeways) 8.48 60.82 Urban multilane undivided highways (nonfreeways) 0.00 1.90 00.014.2sdaorenal-owtnabrU 24.3251.901syaweerflaruR Rural multilane divided highways (nonfreeways) 123.32 73.17 Rural multilane undivided highways (nonfreeways) 0.00 0.61 66.6827.482sdaorenal-owtlaruR 55.75308.925sepytyawdaorllassorcaslatoT Table 10. Total mileage of Minnesota treatment and nontreatment sites considered for inclusion in the safety evaluation of shoulder rumble strips.

46 mation to the best of their capability, but much of the infor- mation remained missing. For those sites where complete information had been provided and the site appeared appro- priate for inclusion in the safety evaluation, and for those sites where installation dates and the construction history were incomplete, the research team turned to the videologs to gather this information. For those candidate treatment sites identified for videolog review, roadway characteristic data were obtained from MoDOT’s Transportation Management System (TMS) database. Using the roadway characteristic data, the candidate treatment sites were grouped according to the eight roadway types of interest for the study. Table 11 shows the total mileage (by roadway type) of treat- ment and nontreatment sites from Missouri considered for inclusion in the safety evaluation. Similar to the previous table, this table reflects total mileage for treatment and non- treatment sites after a series of data quality checks were per- formed to ensure data were consistent and complete for each location. This table does not classify the mileage by installa- tion dates or by total mile-years. This level of detail is pro- vided in the Descriptive Statistics part of this section. Pennsylvania Sites The Pennsylvania DOT (PennDOT) maintains a database of low-cost safety improvements made across the entire state of Pennsylvania. From this database, PennDOT provided a list of approximately 150 safety improvement projects where the installation of shoulder rumble strips was the only safety improvement made as part of the project. This list included the location of the safety projects on the state highway net- work and the installation date of the project. The research team reviewed the locations of these safety projects using PennDOT’s videolog, accessible via the Internet. This review could only confirm that approximately 43 percent of the safety projects from the initial list were actually constructed. In some cases the research team could not confirm the pres- ence of shoulder rumble strips because the quality of the video- log was poor, while in other instances, the quality of the videolog was high, but the videolog showed that milled shoul- der rumble strips were not present at the site. Additional treatment sites were identified either through personal knowledge of the installations or during the process of iden- tifying nontreatment sites for the evaluation. For those treat- ment sites included in the evaluation but not included in the initial list of safety improvement projects provided by Penn- DOT, the research team contacted the PennDOT district offices to inquire about the construction history of the sites. For some districts, PennDOT personnel provided the neces- sary information, while for other districts the research team visited the district offices and reviewed plans, contracts, and other documentation to gather the construction history and installation date data. Comparisons of the dates of the video- logs and the installation dates were used to determine the appropriateness of the site for a before-after evaluation and/or a cross-sectional analysis. Table 12 shows the total mileage (by roadway type) of treatment and nontreatment sites from Pennsylvania consid- ered for inclusion in the safety evaluation. This table is simi- lar to the tables for Minnesota and Missouri. Summary of Sites Across All States Table 13 shows the total mileage (by roadway type) of treat- ment and nontreatment sites summed across all three states (i.e., Minnesota, Missouri, and Pennsylvania) considered for inclusion in the safety evaluation of shoulder rumble strips. Based on the total mileage of both treatment and nontreatment sites across all three states, Table 13 suggests that analyses of the data for urban freeways, rural freeways, rural multilane divided Roadway type Treatment sites (mi) Nontreatment sites (mi) Urban freeways 5.13 0.51 Urban multilane divided highways (nonfreeways) 1.56 6.95 Urban multilane undivided highways (nonfreeways) 0.00 1.92 Urban two-lane roads 0.00 3.15 Rural freeways 77.73 7.91 Rural multilane divided highways (nonfreeways) 21.43 15.04 Rural multilane undivided highways (nonfreeways) 0.00 0.31 Rural two-lane roads 12.85 77.20 Totals across all roadway types 118.70 112.99 Table 11. Total mileage of Missouri treatment and nontreatment sites considered for inclusion in the safety evaluation of shoulder rumble strips.

highways (nonfreeways), and rural two-lane roads have the greatest potential to provide meaningful results. Special attention should be drawn to the total nontreatment miles for rural freeways. During the site selection process it was very difficult to find appropriate nontreatment sites for this particular roadway type. Hundreds of miles of videologs were reviewed in each state to identify sections of rural free- ways without some type of shoulder rumble strip. Because shoulder rumble strips (i.e., milled, rolled, or formed) are pre- dominantly installed on rural freeways in Minnesota, Mis- souri, and Pennsylvania, identifying nontreatment sites for this particular roadway type was extremely difficult. Videolog Data Collection This section describes the videolog data collection effort conducted as part of the safety evaluation of shoulder rumble strips. The videolog review was briefly described above, but this section describes in more detail the types of information gathered during the videolog review. The videolog review served several purposes. First, the videologs were reviewed to confirm the presence or absence of milled shoulder rumble strips at the candidate treatment and nontreatment sites. Second, site characteristic data obtained from roadway inventory databases were verified. Third, roadside data not available from the roadway inven- tory files were collected for each site. Fourth, offset distances of the rumble strips with respect to the edgeline were esti- mated or verified if such information had initially been pro- vided for a site. Finally, the review served to confirm the construction history of a site when possible and to determine how a site should be considered during the analysis. While reviewing sites to confirm the absence or presence of milled shoulder rumble strips, several issues were consid- ered. For treatment sites, the focus of the safety evaluation is on milled shoulder rumble strips, so only those sites with Roadway type Treatment sites (mi) Nontreatment sites (mi) Urban freeways 63.95 19.88 Urban multilane divided highways (nonfreeways) 0.23 2.47 Urban multilane undivided highways (nonfreeways) 0.00 0.46 Urban two-lane roads 1.23 26.88 Rural freeways 70.83 25.23 Rural multilane divided highways (nonfreeways) 5.36 3.43 Rural multilane undivided highways (nonfreeways) 0.00 0.00 Rural two-lane roads 23.34 99.49 Totals across all roadway types 164.94 177.84 Table 12. Total mileage of Pennsylvania treatment and nontreatment sites considered for inclusion in the safety evaluation of shoulder rumble strips. Roadway type Treatment sites (mi) Nontreatment sites (mi) Urban freeways 70.80 131.36 Urban multilane divided highways (nonfreeways) 10.27 70.24 Urban multilane undivided highways (nonfreeways) 0.00 4.28 Urban two-lane roads 3.64 30.03 Rural freeways 257.71 56.56 Rural multilane divided highways (nonfreeways) 150.11 91.64 Rural multilane undivided highways (nonfreeways) 0.00 0.92 Rural two-lane roads 320.91 263.35 Totals across all roadway types 813.44 648.38 Table 13. Total mileage of treatment and nontreatment sites considered for inclusion in the safety evaluation of shoulder rumble strips (includes data from Minnesota, Missouri, and Pennsylvania). 47

milled shoulder rumble strips are included in the safety eval- uation. During the review process, many sites were reviewed where the shoulder contained rolled or formed rumble strips. In some cases, these sites were recorded in the master databases, but these sites are not included in the analyses. Nontreatment sites consist of locations without any type of shoulder rumble strip, whether rolled, formed, or raised. Thus, the analyses compare the crash history of sites with milled shoulder rumble strips to sites without any type of shoulder rumble strip treatment (or any other type of shoul- der treatment) intended to reduce SVROR-type crashes. Sim- ilarly, none of the sites included in the safety evaluation include centerline rumble strips. When confirming the pres- ence of the milled shoulder rumble strip at a site, the begin- ning and ending locations of the milled shoulder rumble strips were recorded. For divided highways (i.e., freeways and multilane divided highways [nonfreeways]), a site was con- sidered a valid treatment site when rumble strips were located on the right (outside) shoulder. When possible, the presence of shoulder rumble strips installed on the left (inside) shoul- der was recorded as well. Most of the divided highways had rumble strips on both the right (outside) and left (median) shoulders. The videologs were used to verify certain roadway charac- teristic data that were considered potentially important in explaining the safety effectiveness of shoulder rumble strips. Prior to the videolog data collection, roadway inventory data files were obtained for the three states involved in the safety evaluation. The roadway characteristic data verified during the videolog review included the following: • Area type (rural vs. urban); • Roadway type (i.e., freeway, multilane divided, multilane undivided, or two-lane); • Number of lanes; • Lane widths; and • Shoulder widths. During the videolog review, the roadside hazard rating (RHR) was recorded for both sides of the roadway. The RHR system was developed by Zegeer et al. (66) to characterize the accident potential for roadside designs found on two-lane roads. The roadside hazard is ranked on a 7-point categorical scale from 1 (safest and most traversable) to 7 (most danger- ous and least traversable). For undivided roadways, the RHR was recorded separately for the right and left sides of the road, and for divided highways, the RHR was recorded separately for the right and left (median) sides of the road in both direc- tions of travel. For Minnesota, which treats the separate directions of travel of divided highways as a single site, this meant that four RHRs were recorded for each site. For Mis- souri and Pennsylvania, which treat the separate directions of travel of divided highways as separate sites, this meant that two RHRs were recorded for each site on a divided highway. The RHR system is a subjective measure for quantitatively characterizing the accident potential of the roadside. Several data collectors could view a given site and, based on their sub- jective opinion, assign a different RHR to the site. To mini- mize the variability between data collectors in assigning RHRs, several steps were taken. First, written and pictorial descrip- tions of the seven roadside hazard categories that distin- guished the differences between categories were provided to each data collector involved in the study. These written and pictorial descriptions, provided in Appendix D, were taken from an FHWA report on the expected safety performance of two-lane roads (67) and from a website for the Interactive Highway Safety Design Model (IHSDM), but as noted above, the original categories/descriptions were developed by Zegeer et al. (66). Next, 10 sites were chosen for a pilot study in which each data collector independently reviewed the sites and assigned an RHR to each. Then the variability in ratings between data collectors was evaluated. In most cases, the dif- ferences in ratings were plus or minus 1 or 2. Then, 5 of the original 10 sites with the greatest variability in assigned RHR among reviewers were examined by the data collectors as a group to discuss what each data collector was considering when rating each site. During this group discussion, several members with the most experience in using the roadside haz- ard rating explained their thought process and how they would rate the sites. This pilot effort could not completely eliminate the variability between data collectors in assigning roadside hazard rating scores, but the exercise was designed to minimize that variability. When reviewing the treatment sites, the offset distances of the rumble strips from the edgeline were estimated to the nearest inch. Estimating the offset distances from videologs was not the most accurate method of measurement, but resource limitations prohibited visiting each site in the field, which would have been the best approach for obtaining the most accurate measurements. The offset distances were judged taking into account relative distances of the travel lane widths, shoulder widths, the width of the rumble strips, and the width of the edgeline. The offset distance was measured relative to the inside edge of the edgeline separating the out- side travel lane from the shoulder so, for example, if the rum- ble strip was installed on the edgeline, then the offset distance would be 0 in. (0 mm), while if the rumble strip was installed adjacent to the edgeline but not on the edgeline, then the off- set distance was estimated to be 4 in. (102 mm) because most standard edgelines are 4 in. (102 mm) in width. Offset dis- tances were only verified and/or estimated for rumble strips on the right (outside) shoulder. No attempt was made to esti- mate or verify the accuracy of the offset distance for rumble strips on the left shoulder because the perspectives of the 48

videologs of the left (median) shoulders of divided highways made it very difficult to estimate these offset distances. The videolog review also served to gather information con- cerning the construction history of sites and installation dates, which in turn was useful in determining the appropriateness of sites for a before-after EB analysis and/or a cross-sectional analysis and in determining analysis periods for individual sites. The dates of the videologs were compared to the instal- lation dates to verify the accuracy of the data and/or establish new installation dates for the rumble strips. The videologs in some cases also revealed when sites were under construction, which was taken into consideration when determining the analysis periods for sites or served as justification for exclud- ing a site from the analysis because there was some uncertainty about the construction history and potentially the accuracy of the roadway characteristic data. When possible, the videologs were used to assess the conditions of treatment sites prior to installation of the rumble strips. This information was useful for determining whether a treatment site could be included in an EB analysis using before-after data or to include the site only in a cross-sectional analysis because the data collector noticed something from the videolog that created uncertainty about the condition of the treatment site prior to installation of the milled shoulder rumble strips. For example, if adjacent segments upstream or downstream of a treatment site had formed rumble strips that appeared to have been present for years, but the treatment site with milled shoulder rumble strips had a relatively recent installation date, it is very possi- ble that the conditions of the treatment site prior to installa- tion of the milled shoulder rumble strips included formed shoulder rumble strips. In this instance, it would not be appropriate to include the treatment site in a before-after EB analysis because the before conditions likely included formed rumble strips on the shoulders. It would, however, be perfectly acceptable to include it as a treatment site in a cross-sectional analysis but consider only those years since installation of the milled shoulder rumble strips. To identify nontreatment sites for possible inclusion in the safety evaluation, certain cross-sectional characteristics (e.g., lane widths and shoulder widths) and traffic volumes of each potential treatment site were reviewed. For those roadway types for which a reasonable amount of mileage of treatment locations had been identified, a list of locations was generated for review using roadway characteristics data. An attempt was made to generate a list of nontreatment locations for each roadway type with a range of lane widths, shoulder widths, and traffic volumes that matched the ranges of the respective characteristics of the treatment sites. This list was generated without knowledge of whether rumble strips were present at the respective locations. Therefore, the videologs were used to review these locations for the presence/absence of shoul- der rumble strips. For certain roadway types, limited mileage of nontreatment sites were found after reviewing this initial list. In those instances, maps were used to identify selected routes and extensive lengths of roadway mileage were reviewed in an attempt to identify nontreatment locations for inclusion in the analysis. In other situations, nontreatment locations were identified while reviewing the videologs for treatment sites. When a potential treatment site was reviewed but no rumble strips were present at the sites, the appropriate data were collected for the site, and the site was classified as a non- treatment rather than a treatment site. The following final points regarding the data collection effort, relevant to the analysis approach and analysis results, are noteworthy: • For divided highways when rumble strips are present on the left (median) shoulder, it is always assumed that these rumble strips were installed during the same calendar year as the rumble strips on the right (outside) shoulder. • Even when milled rumbles strips are installed continuously along a segment, there are many breaks in the rumble strips for various reasons such as bridges, speed-change lanes (i.e., deceleration and acceleration lanes), intersections, driveways, etc. Depending upon the roadway type and the policy of the individual states, the frequencies of these breaks vary considerably. For example, in some cases, shoulder rumble strips near interchanges are installed along the shoulder of a speed-change lane, and in other instances the rumble strips are dropped/begin at the beginning/end of the speed-change lane. Also the frequency/spacing of inter- changes differs depending upon whether it is an urban or rural area. For those treatment sites where a significant length of contiguous mileage of shoulder rumble strip installations existed, long breaks in the rumble strips may have been recorded during the data collection process, but the boundaries of the treatment site were not modified to reflect the breaks in the rumble strips. Thus, there are loca- tions along the roadways of treatments sites where shoul- der rumble strips are not present. Only those treatment sites where shoulder rumble strips were not installed over a very long stretch of highway were the boundaries of the treatment sites modified to reflect numerous or significant breaks in the continuous shoulder rumble strips. Based purely on observation, the urban treatment sites tended to have more natural breaks within the sites compared to the rural treatment sites. To further explain how the data were collected, the following examples are provided: – Example 1: The treatment site is a 10 mi (16 km) segment along a rural freeway, and three interchanges occur within this 10 mi (16 km) segment. At each interchange, the shoulder rumble strips are discontinued at the beginning of the deceleration lane, are installed between the gore points, and continue after the end of the acceleration 49

lane. Even though the rumble strips are not installed along the full 10 mi (16 km) segment, the full 10 mi (16 km) seg- ment would have been recorded as a treatment site. – Example 2: The treatment site is initially listed as a 0.5 mi (8 km) segment of rural two-lane road, but on one end of the treatment site there is a long bridge so the rumble strips were only installed over a 0.4 mi (0.64 km) segment of highway. In this case, the boundaries of the treatment site would have been defined to be 0.4 mi (0.64 km) in length. • The ideal type of treatment site to include in a before-after analysis is one in which the only type of treatment made during the analysis period is the safety improvement (i.e., installation of shoulder rumble strips). For many of the treatment sites in Minnesota and Missouri, the shoulder rumble strips were installed as part of a resurfacing project. To the best of our ability, we obtained information through various means to confirm that the cross-sectional charac- teristics of the roadway did not change (i.e., number of lanes, lane widths, and shoulder widths). For those treat- ment sites where we could confirm that the cross-sectional characteristics did not change, the site was identified as being appropriate for use in a before-after analysis. If it was determined that the cross-sectional characteristics changed as a result of the resurfacing, then the treatment site was identified as being appropriate for a cross-sectional analysis. As a result of this decision, the safety effectiveness of shoul- der rumble strips will be confounded to some degree with the safety effects of resurfacing a roadway. • In Pennsylvania, the initial list of treatment sites was gen- erated from a low-cost safety improvement database devel- oped and maintained by PennDOT. This database includes information such as the project number, type of safety improvement, installation date, location of the improve- ment, etc. This database is an inventory of the low-cost safety improvements made by PennDOT throughout the entire state roadway network and includes many types of low-cost safety improvements, not just the installation of shoulder rumble strips. It is our understanding that Penn- DOT has a process for identifying high-crash locations, and through this process PennDOT programs the imple- mentation of certain types of low-cost safety improvements such as shoulder rumble strips. Thus, the locations of some of the treatment sites in Pennsylvania were initially identi- fied as being high-crash locations compared to the rest of the highway network. For the other treatment sites included in the safety evaluation but not initially identified through the low-cost safety improvement database, the rumble strips may or may not have been installed as part of a broader proactive safety policy to install shoulder rumble strips on certain types of roadways. For Minnesota and Missouri, the policy for determining the need and location for instal- lation of shoulder rumble strips is not known. It is likely that for the nonfreeway roadways, each state has a procedure/ program for identifying high-crash locations (e.g., loca- tions with high frequencies of SVROR crashes) and that some of the rumble strip installations being analyzed as part of this evaluation were implemented as part of such a program. • All milled rumble strips are treated as being equivalent in their alerting properties. Although an effort was made to obtain rumble strip dimensions for each treatment site, this information was very difficult to obtain, and in many cases when it was obtained, the validity of the data was questionable. Therefore, in the analysis, a site with rumble strip dimensions of 16, 6, 0.5, and 12 in. (406, 152, 13, and 305 mm) for the length, width, depth, and spacing would be treated the same as a site with rumble strip dimensions of 6, 5, 0.375, and 16 in. (152, 127, 10, and 406 mm) for the length, width, depth, and spacing. Database Development The final database(s) utilized for analysis consisted of the roadway characteristic data (including traffic volume), the pri- mary data from the videolog data collection effort, and crash data. In summary, the database(s) for each state included the following roadway inventory and videolog data for a given site: • Location reference information (i.e., beginning and ending mileposts/logpoints, or route, county, segment, and offsets); • Area type (rural vs. urban); • Roadway type (i.e., freeway, multilane divided, multilane undivided, or two-lane); • Number of lanes; • Lane widths; • Shoulder widths; • Presence/absence of milled shoulder rumble strips; • Offset distance of rumble strips from edgeline; • RHR; • Analysis period(s) (including year(s) without rumble strips, installation year(s), and years with rumble strips); and • ADT for each year in the analysis period(s). Concerning the traffic volume data, the original roadway inventory files obtained from the states did not contain ADTs for all sites for each year in the analysis period(s). Therefore, rules were established for interpolating and extrapolating the ADT data so that the final database included ADTs for each site for each year of the analysis period(s). The analysis period(s) were determined based upon the construction his- tory and installation data gathered and the years of available crash data for each state. Crash data were obtained for the fol- lowing calendar years (inclusive) for each state: 50

• Minnesota (1997 through 2005), • Missouri (1997 through 2006), and • Pennsylvania (1997 through 2006). The types of crash data in the final database(s) for potential use in the analyses consisted of the following: • Crash report number or crash ID number, • Date of crash, • Location information (county, route, direction, segment and offset or logpoint), • Number of vehicles involved, • Crash severity, • Accident type or manner of collision, • Run-off-road indicator, • Fatigue-related indicator, • Heavy vehicle indicator, • Adverse pavement indicator, • Light condition indicator, and • Alcohol/drug indicator. The final database only included crashes assigned to road- way segments. Rules were established to eliminate (i.e., screen out) intersection-related crashes and crashes near inter- changes that did not occur on or adjacent to the mainline freeway. For example, crashes that occurred on interchange ramps (i.e., the ramp proper) are not included in the data- base, while crashes that occurred within or adjacent to accel- eration or deceleration lanes are included in the database. Several other rules were established for developing the final database(s). Most of these rules pertained to establishing a rationale for combining adjacent sites to create longer homo- geneous sites. Several of these rules pertained to: • Selected roadway characteristic (e.g., lane widths, shoulder widths, number of lanes); • ADT thresholds; and • Desirable minimum lengths (e.g., 0.3 mi [0.48 km]). The following section provides descriptive statistics of the information contained in the databases developed for the safety evaluation. Descriptive Statistics The basic study layout and descriptive statistics in either tab- ular and/or graphical form are provided for the independent variables (i.e., ADT, site geometrics) and dependent variables (i.e., crash data) of interest in the safety evaluation of shoulder rumble strips. General study layout. Data at each site were collected over periods of varying lengths (i.e., number of years). For compar- ison, the site length and the number of years were combined into a single variable, mile-years, for each site. Throughout the remainder of this section on the safety evaluation of shoulder rumble strips, the four site types are encoded as follows for ease of readability: • BA-No RS: Nontreatment site of the matched before-after site pair in the before period; • BA-RS: Treatment site of the matched before-after site pair in the after period; • CS-No RS: Nontreatment cross-sectional site; and • CS-RS: Treatment cross-sectional site. Table 14 summarizes the basic layout of the available data in the three states, separately for each roadway type and type of site: number of sites, total site length, and mile-years. Because of insufficient number of sites and mile-years or lack of comparison sites for a number of roadway types and states to conduct the safety evaluation, it was decided to focus the safety evaluation of shoulder rumble strips on the following four categories: • Urban freeways in Pennsylvania only; • Rural freeways in Missouri and Pennsylvania only; • Rural multilane divided highways (nonfreeways) in all three states; and • Rural two-lane roads in all three states. The analysis for rural freeways does not include Min- nesota data even though there appears to be a sufficient number of mile-years of cross-sectional sites with and with- out rumble strips for analysis purposes; however, rural free- ways in Minnesota were not included in the analysis because the distribution of ADT was unbalanced between treatment and nontreatment sites. Many of the Minnesota sites with rumble strips had lower ADTs than the Minnesota sites without rum- ble strips. This unbalanced distribution occurred primarily due to the difficulty of finding rural freeway nontreatment sites (see earlier part on Site Selection in this Section). ADT volume. For each site, ADTs were first averaged across years within an analysis period. This allowed for a fair comparison of the distribution of ADTs across site types, analysis periods, and states since the sample size is reduced to the number of sites within each category and thus not unduly influenced by the length of the varying analysis periods. Figures 7 through 10 show the ADT distributions in the form of side-by-side boxplots, separately for each of the four roadway types discussed above: urban freeways, rural free- ways, rural multilane divided highways (nonfreeways), and rural two-lane roads. Within each figure, the data are orga- nized by state when more than one state is included in the analysis; within each state, the data are ordered by site type— 51

Minnesota Missouri Pennsylvania Roadway type Site type Number of sites Length (mi) Mile- years Number of sites Length (mi) Mile- years Number of sites Length (mi) Mile- years 03.63SRoN-AB 182.00 BA-RS 0 6 5.13 9.87 53 35.51 129.60 CS-No RS 49 110.97 998.75 2 0.51 5.07 37 19.88 194.50 Urban freeways CS-RS 1 1.72 8.61 0 48 28.44 104.00 31.148.908.33SRoN-AB BA-RS 3 5.60 10.98 4 1.56 4.20 1 0.23 0.91 CS-No RS 30 60.82 547.34 7 6.95 69.50 7 2.47 23.64 Urban multilane divided highways (nonfreeways) CS-RS 3 2.88 13.49 0 0 SRoN-AB BA-RS 0 0 0 CS-No RS 2 1.90 17.08 1 1.92 19.20 1 0.46 4.58 Urban multilane undivided highways (nonfreeways) CS-RS 0 0 0 19.699.6SRoN-AB BA-RS 1 1.40 4.19 0 3 1.23 3.69 CS-No RS 0 3 3.15 31.50 29 26.88 261.60 Urban two-lane roads CS-RS 1 1.01 5.06 0 0 SRoN-AB 313.00 80.09 BA-RS 0 29 52.08 156.00 18 15.43 57.02 CS-No RS 8 23.42 210.75 6 7.91 79.20 16 25.23 245.60 Rural freeways CS-RS 28 109.15 495.67 12 25.65 77.90 41 55.40 146.30 BA-No RS 100.00 109.00 32.57 BA-RS 6 20.00 60.00 14 19.11 51.20 5 4.18 4.61 CS-No RS 27 73.17 658.49 12 15.04 150.00 8 3.43 32.88 Rural multilane divided highways (nonfreeways) CS-RS 27 103.32 508.56 1 2.32 2.32 4 1.18 2.35 SRoN-AB BA-RS 0 0 0 CS-No RS 1 0.61 5.46 1 0.31 3.13 0 Rural multilane undivided highways (nonfreeways) CS-RS 0 0 0 BA-No RS 478.41 64.00 136.20 BA-RS 28 95.51 285.67 5 10.52 30.70 20 23.34 69.90 CS-No RS 28 86.66 776.29 32 77.20 772.00 90 99.49 933.20 Rural two-lane roads CS-RS 53 179.21 851.60 1 2.33 2.33 0 a Shaded cells are the focus of statistical analysis. Table 14. Summary study layout—total number of sites, site length, and mile-years by state, roadway type, and site typea.

before-after sites then cross-sectional sites. The mean ADTs of nontreatment sites are colored white; those of the treat- ment sites are black. Each figure also contains a table of basic descriptive ADT statistics: number of sites, mean, standard deviation, minimum, median, and maximum. Lane width. Lane widths ranged from 7 to 20 ft (2.1 to 6.1 m) across all sites and states, with the majority of lanes being 12 ft (3.6 m) wide. The distribution of lane width is summarized in Table 15 by state and site type. Due to the lack of variability in lane width, it was decided to exclude this vari- able from all modeling efforts. Outside and inside shoulder widths. Outside shoulder widths ranged from 1 to 14 ft (0.3 to 4.3 m) across sites and states, with slightly over half of outside shoulders being 10 ft (3.0 m) wide. The distribution of outside shoulder width is summarized in Table 16 by state and site type. For divided highway sites with an inside shoulder, inside shoulder widths ranged from 1 to 10 ft (0.3 to 3.0 m) across sites and states, with approximately 61 percent of inside shoulders being 4 ft (1.2 m) wide. The distribution of inside shoulder width is summarized in Table 17 by state and site type. Due to the lack of variability in either shoulder widths within a given roadway type and/or due to a high correlation with roadside hazard ratings, it was decided to exclude shoulder width from all modeling efforts. The use of RHRs discussed below addresses how shoulder width information is being captured in the statistical models. RHR. Outside and inside (on divided highways) road- side hazard ratings were recorded as integers ranging from 1 (low RHR) to 7 (high RHR); both variables are treated as continuous variables in the statistical model development. Tables 18 (outside RHR) and 19 (inside RHR) present basic descriptive RHR statistics—number of sites and minimum, maximum, mean, and standard deviation—across sites within 53 Roadway type=1:Urban freeways trmt_status No RS RS 0 20,000 40,000 60,000 80,000 100,000 AD T (av g ve hs /d a y) 0 boxes cl ipped PA BA CS 35No. of sites 701,62naeM 434,21veddtS 452,11niM 715,22naideM 290,95xaM 53 28,549 12,553 13,830 25,659 59,391 37 51,791 18,400 11,856 54,982 92,757 48 43,493 20,725 13,785 41,283 91,223 Figure 7. ADT distribution by site type for urban freeways in Pennsylvania.

a given roadway type, state, and site type. Since RHR is related to shoulder width, these two explanatory variables are not statistically independent. Including RHR in the modeling effort and excluding shoulder width therefore solves the issue of non-independent variables; also, RHR accounts for addi- tional variability in roadside factors that shoulder width does not. Non-integer values for minimums and maximums in Tables 18 and 19 are the result of combining adjacent seg- ments into homogeneous sites for analysis purposes. When adjacent segments with different RHRs were combined into a single site for analysis purposes, a weighted average, based on segment length, of the RHR was calculated for the site. Rumble strip offset. Rumble strip offset, measured in inches, is available for treatment sites only. A preliminary check of offset measurements on a continuous scale led to considering this variable as a categorical variable in two ways for the statistical analysis: • Two categorical levels: – Edgeline rumble strips (i.e., offset distances of 0 to 8 in. [0 to 203 mm]), and – Non-edgeline rumble strips (i.e., offset distances of 9+ in. [229+ mm]). • Three categorical levels: – 0 to 8 in. (0 to 203 mm), – 9 to 20 in. (229 to 508 mm), and – 21+ in. (533+ mm). The distribution of rumble strip placement across the two offset levels is shown in Columns 3 and 4 of Table 20 by road- way type and state. The distribution of rumble strip offset across the three offset levels is shown in the last two columns of Table 20. Overall, just over half (56 percent) of the offset distances are in the 9 to 20 in. (229 to 508 mm) range; another 32 percent are in the 0 to 8 in. (0 to 203 mm) range; and the remaining 12 percent are in the 21+ in. (533+ mm) range. 54 Roadway type=5:Rural freeways trmt_status No RS RS 0 10,000 20,000 30,000 40,000 AD T (av g ve hs /d ay ) 0 boxes cl ipped MO PA BA CS BA CS No. of sites Mean 25,118 Std dev 5,464 Min 11,539 Median 25,752 Max 36,643 29 26,450 6,212 13,079 28,702 37,112 6 22,634 5,883 16,870 21,370 29,410 12 22,013 8,064 12,438 24,123 32,354 18 13,091 4,012 6,963 11,795 24,752 18 13,615 3,602 6,777 14,584 21,755 16 15,680 4,049 7,192 15,209 22,571 41 18,572 8,010 10,319 15,930 34,406 29 Figure 8. ADT distribution by site type for rural freeways in Missouri and Pennsylvania.

Recovery area. Recovery area at each treatment site was calculated as the difference between shoulder width and rum- ble strip offset. The final measurement is in feet. Based on the distribution of this variable, recovery area was treated as a cat- egorical variable at two levels in the statistical models: 0 to 4 ft (0 to 1.2 m) and over 4 ft (1.2+ m). For nontreatment sites, the recovery area equals the shoulder width; therefore shoul- der width was categorized in the same fashion: 0 to 4 ft (0 to 1.2 m) and over 4 ft (1.2+ m). Since recovery area was only used in conjunction with rumble strip offset in the statistical modeling to account for shoulder width, Table 21 presents the distribution of recovery area within each type of rumble strip placement (i.e., edgeline or non-edgeline). For nontreat- ment sites (shaded rows in Table 21), the table presents the distribution of shoulder width across the two levels. In sum- mary, six combinations of rumble strip offset and recovery areas were considered: 1. RS edgeline and 4+ ft RA, 2. RS edgeline and 0–4 ft RA, 3. RS non-edgeline and 4+ ft RA, 4. RS non-edgeline and 0–4 ft RA, 5. No RS and 4+ ft shoulder width, and 6. No RS and 0–4 ft shoulder width. Crash data. Four crash types are considered in the safety evaluation of shoulder rumble strips: 1. TOT crashes, 2. FI crashes, 3. SVROR crashes, and 4. SVROR FI crashes. An SVROR crash was defined to be any single-vehicle crash that involved a vehicle leaving the travel way that was not inter- section or ramp related. 55 Roadway type=6:Rural multilane divided highways (nonfreeways) trmt_status No RS RS 0 5,000 10,000 15,000 20,000 25,000 30,000 35,000 AD T (av g ve hs /d a y) 0 boxes cl ipped MN MO PA BA CS BA CS BA CS No. of sites Mean 6,193 Std dev 626 Min 5,110 Median 6,189 Max 6,875 66 6,292 850 4,959 6,337 7,459 27 15,827 6,036 8,349 14,957 31,692 27 14,316 5,144 6,005 13,383 22,665 14 11,781 4,143 6,192 11,935 20,763 14 11,945 3,205 5,326 12,906 15,198 12 11,074 3,946 4,956 11,825 15,725 1 5,725 . 5,725 5,725 5,725 5 13,041 3,250 10,587 12,009 18,753 5 11,185 1,391 9,653 11,424 13,195 8 13,421 2,956 8,267 13,306 17,018 4 15,876 2,551 12,874 16,269 18,093 Figure 9. ADT distribution by site type for rural multilane divided highways (nonfreeways) in Minnesota, Missouri, and Pennsylvania.

56 Roadway type=8:Rural two-lane roads trmt_status No RS RS 0 2,500 5,000 7,500 10,000 12,500 15,000 AD T (av g ve hs /d a y) 0 boxes cl ipped MN MO PA BA CS BA CS BA CS No. of sites Mean 3,736 Std dev 2,162 Min 782 Median 3,216 Max 9,288 2828 4,081 2,466 921 3,334 10,386 28 3,040 2,822 180 1,471 8,981 53 3,563 1,446 1,285 3,633 7,431 5 3,206 2,656 861 2,203 6,077 5 3,087 2,536 983 1,821 6,205 32 5,211 3,091 952 4,681 12,776 1 6,348 . 6,348 6,348 6,348 20 4,642 2,086 1,081 4,111 9,067 20 4,299 2,373 948 3,734 8,674 90 4,445 2,319 910 4,422 10,177 Figure 10. ADT distribution by site type for rural two-lane roads in Minnesota, Missouri, and Pennsylvania. ainavlysnnePiruossiMatosenniM Cross-sectional sites Cross-sectional sites Cross-sectional sites Lane width (ft) Before– after sites Untreated Treated Before– after sites Untreated Treated Before– after sites Untreated Treated 7 1 10 2 14 9 11 1 8 1 4 12 50 12 31 44 76 46 32 14 83 78 90 13 1 2 1 5 14+ 1 1 2 1 8 3 Table 15. Distribution of lane width by state and site type. Analyses of TOT crashes are performed primarily because several previous safety evaluations of shoulder rumble strips analyzed this crash type. However, analyses of TOT crashes include many other crash types besides SVROR crashes (i.e., the target crash type). No strong argument can be made to support why shoulder rumble strips would affect crashes other than SVROR crashes. Analyses based on FI crashes were also conducted because there is great interest in reduc- ing crashes that result in fatalities and injuries, but again, analyses of FI crashes include many other crash types besides the target crashes. Analyses of SVROR crashes are expected to produce more reliable results than analyses of TOT and FI

57 ainavlysnnePiruossiMatosenniM Cross-sectional sites Cross-sectional sites Cross-sectional sites Outside shoulder width (ft) Before– after sites Untreated Treated Before– after sites Untreated Treated Before– after sites Untreated Treated 0 10 1 3 2 2 1 2 11 3 1 7 10 4 5 7 6 1 2 10 32 1 5 1 3 1 10 6 1 7 4 3 9 2 10 7 1 5 7 4 1 8 8 16 22 3 3 10 15 6 9 3 4 3 4 1 10 15 11 35 40 24 14 72 44 77 11 1 3 4 12 2 14 1 Table 16. Distribution of outside shoulder width by state and site type. iruossiMatosenniM a ainavlysnneP Cross-sectional sites Cross-sectional sites Cross-sectional sites Inside shoulder width (ft) Before– after sites Untreated Treated Before– after sites Untreated Treated Before– after sites Untreated Treated 0 2 2 4 14 7 1 1 4 2 2 1 1 2 3 3 6 15 19 3 1 4 7 4 60 35 70 5 1 1 7 1 1 6 2 2 3 7 1 8 1 9 2 10 2 3 a Inside shoulder width data not available for Missouri sites. Table 17. Distribution of inside shoulder width by state and site type. crashes because the analyses include only those crashes expected to be most directly impacted by shoulder rumble strips. Finally, analyses based on SVROR FI crashes are of interest because these analyses address the more severe tar- get crashes. The crash data across all years are summarized in Table 22 and are shown as both total number of crashes and crash fre- quency (crashes/mi/yr), separately for each type of site of a given roadway type within a given state. The two statistics are organized within each crash type in the following order: TOT, FI, SVROR, and SVROR FI crashes. This breakdown of the data is the level at which the statistical analyses are performed. Table 22 also provides the number of sites and their total length and mile-years to facilitate comparison between groups of data. For before-after site pairs (i.e., same site paired in time), number of sites and length are shown only once since the sites are the same before and after treatment; however, since the study periods changed from site to site, mile-years vary between nontreatment and treatment before- after site pairs. The crash data are summarized by roadway type, state, site type, and rumble strip position (edgeline, non-edgeline, and no rumble strips) in Table 23. The crash count and frequency for TOT, FI, SVROR, and SVROR FI are presented. The crash summaries for the supplemental analyses con- ducted for heavy vehicle, adverse pavement condition, and low-light condition crashes are presented in Table 24. The data are for SVROR crashes only and are summarized by road- way type, state, site type, and treatment status. Adverse pave- ment condition crashes are defined as crashes that occurred under wet, snow, slush, ice, standing or moving water, muddy, debris, or oily road surface conditions. Low-light condition crashes are defined as crashes that occurred during dawn, dusk, or dark.

Analysis Approach The safety evaluation of shoulder rumble strips is based on the comparison of crash frequencies between treatment and nontreatment sites. This comparison is made separately for each combination of crash type—TOT, FI, SVROR, and SVROR FI crashes—and roadway type of interest. Compar- isons of the crash frequencies are made separately for each state and across states for three of the four roadway types of interest. The following two statistical approaches are used to evaluate whether installing shoulder rumble strips has an effect on crash frequencies: 58 Outside RHR Roadway type State Site type Number of sites Minimum Maximum Mean Standard deviation BA 53 1.0 4.0 3.3 0.7 CS-No RS 37 1.0 5.0 3.6 1.0 Urban freeways PA CS-RS 48 2.7 4.0 3.6 0.4 BA 29 2.0 4.0 3.0 0.5 CS-No RS 6 3.0 3.7 3.2 0.3 MO CS-RS 12 2.0 4.0 3.0 0.4 BA 18 2.0 4.0 3.4 0.6 CS-No RS 16 2.0 4.0 3.2 0.7 Rural freeways PA CS-RS 41 2.2 4.1 3.6 0.5 BA 6 2.0 2.0 2.0 0.0 CS-No RS 27 2.0 4.0 2.9 0.9 MN CS-RS 27 1.0 3.3 2.2 0.6 BA 14 2.0 3.1 2.9 0.4 CS-No RS 12 2.0 4.0 3.0 0.7 MO CS-RS 1 3.0 3.0 3.0 BA 5 3.0 4.0 3.5 0.5 CS-No RS 8 2.8 4.4 3.5 0.6 Rural multilane divided highways (nonfreeways) PA CS-RS 4 3.0 4.0 3.3 0.5 BA 28 1.0 3.0 2.0 0.6 CS-No RS 28 1.0 5.0 2.6 1.4 MN CS-RS 53 1.0 4.0 2.3 0.8 BA 5 3.0 4.0 3.6 0.4 CS-No RS 32 3.0 5.5 3.8 0.7 MO CS-RS 1 3.6 3.6 3.6 BA 20 2.5 5.0 3.5 0.7 Rural two-lane roads PA CS-No RS 90 1.7 6.0 4.1 1.0 Table 18. Outside RHR statistics by roadway type, state, and site type. Inside RHR Roadway type State Site type Number of sites Minimum Maximum Mean Standard deviation BA 53 1.0 5.0 2.5 0.9 CS-No RS 37 2.0 5.0 3.7 1.2 Urban freeways PA CS-RS 48 1.0 5.3 3.3 1.2 BA 29 1.0 4.1 3.3 1.0 CS-No RS 6 2.0 5.0 3.4 1.3 MO CS-RS 12 2.0 4.0 3.0 1.0 BA 18 1.0 5.0 3.1 1.3 CS-No RS 16 2.0 4.0 2.6 0.7 Rural freeways PA CS-RS 41 1.0 5.0 3.6 1.0 BA 6 2.0 2.0 2.0 0.0 CS-No RS 27 2.0 4.0 2.9 0.9 MN CS-RS 27 1.0 3.4 2.2 0.6 BA 14 2.0 3.4 2.5 0.5 CS-No RS 12 0.5 3.2 2.4 0.9 MO CS-RS 1 3.0 3.0 3.0 BA 5 4.0 5.0 4.4 0.5 CS-No RS 8 3.0 5.0 4.2 0.7 Rural multilane divided highways (nonfreeways) PA CS-RS 4 3.0 3.8 3.6 0.4 Table 19. Inside RHR statistics by roadway type, state, and site type.

• A before-after comparison using the EB method applied to the before-after sites, and • A cross-sectional analysis using a generalized linear model (GLM) approach based on crash data from: – all treatment and nontreatment sites (i.e., before-after and cross-sectional sites); and – all before-after sites and all cross-sectional nontreatment sites (i.e., cross-sectional treatment sites are excluded). The rationale, differences, and similarities of the two meth- ods are discussed next. Before-After EB Analysis to Determine the Safety Effectiveness of Shoulder Rumble Strips on Different Roadway Types The EB method is now the most widely used method to evaluate the safety effectiveness of a countermeasure given a set of matched before-after sites and a set of reference sites. The EB method, which adjusts for the effects of regression to the mean, is based on the comparison of observed crash fre- quencies in the after period to predicted crash frequencies in the after period had the treatment not been implemented. To implement the EB methodology, it is crucial to develop a safety performance function (SPF) for each crash type on a particular roadway type based on crash data from a set of nontreatment reference sites. The EB method used in this analysis is the method used in the countermeasure evaluation tool of the FHWA Safety Analyst software (68). The EB methodology is described in a white paper available on the SafetyAnalyst web site (www.safetyanalyst.org); a revised white paper presenting the EB methodology is currently under development (69). The EB methodology is based on methods recommended by Hauer (70) and Hauer et al. (71). The sequence of steps in applying the EB methodology is as follows: • Obtain data for the observed crash frequency on each treatment site during both the before and after periods. • Using the reference group data (i.e., sites that were not improved during the study period) for the entire period during which data are available, develop SPFs that model crash frequencies as a function of site parameters (e.g., traf- fic volumes and site geometrics). This is generally done by means of negative binomial (NB) regression analysis. • Estimate the predicted crash frequency at each treatment site during the before period using the SPF developed for that type of site. • Compute a weighted-average of the predicted and observed crash frequencies at each treatment site during the before 59 Roadway type State Rumble strip placement Number of sites Offset (in.) Number of sites Edgeline 0 0-8 0 9-20 101 APsyaweerfnabrU Non-edgeline 101 21+ 0 Edgeline 24 0-8 24 9-20 2 MO Non-edgeline 17 21+ 15 Edgeline 4 0-8 4 9-20 51 Rural freeways PA Non-edgeline 55 21+ 4 Edgeline 7 0-8 7 9-20 20 MN Non-edgeline 26 21+ 6 Edgeline 5 0-8 5 9-20 0 MO Non-edgeline 10 21+ 10 Edgeline 1 0-8 1 9-20 7 Rural multilane divided highways (nonfreeways) PA Non-edgeline 8 21+ 1 Edgeline 61 0-8 61 9-20 17 MN Non-edgeline 20 21+ 3 Edgeline 1 0-8 1 9-20 0 MO Non-edgeline 5 21+ 5 Edgeline 15 0-8 15 9-20 5 Rural two-lane roads PA Non-edgeline 5 21+ 0 Table 20. Distribution of rumble strip placement and offset by roadway type and state.

period. This crash frequency is referred to as the EB- adjusted expected crash frequency. • Using the EB-adjusted expected crash frequency at each site during the before period, make an estimate of the expected crash frequency at each treatment site during the after period had no change been made. This step of the analysis accounts for changes in traffic volumes between the before and after periods. • Compare the observed after crash frequencies at the treat- ment sites to the expected after crash frequencies at the treat- ment sites had the change not been made. The difference between these observed and expected crash frequencies is an estimate of the safety effectiveness of the treatment. SPFs were developed for each crash type, roadway type, and state considered based on all nontreatment sites, that is, all before sites and all nontreatment cross-sectional sites. The deci- sion to include the before sites into the reference group was made to use the maximum number of sites and thus to capital- ize on the maximum amount of information to develop the functions. Since, as evidenced by the crash rates, the treatment sites were unlikely to be selected based on a high crash count in thebeforeperiod,this approach, on balance, seemed reasonable. Of the independent variables summarized in the previous section, an attempt was made to incorporate as many variables as possible in the SPF, in addition to ADT, to obtain the best possible function to predict crashes at sites without shoulder 60 Roadway type State Rumble strip placement Recovery area (ft) or shoulder with (ft)a Number of sites 4+ 100 Non-edgeline 0-4 1 4+ 82 Urban freeways PA No rumble strips 0-4 8 Edgeline 4+ 24 Non-edgeline 4+ 17 MO No rumble strips 4+ 35 Edgeline 4+ 4 4+ 53 Non-edgeline 0-4 2 4+ 32 Rural freeways PA No rumble strips 0-4 2 Edgeline 4+ 7 Non-edgeline 4+ 26 MN No rumble strips 4+ 33 4+ 4 Edgeline 0-4 1 Non-edgeline 4+ 10 4+ 25 MO No rumble strips 0-4 1 Edgeline 4+ 1 Non-edgeline 4+ 8 4+ 12 Rural multilane divided Highways (nonfreeways) PA No rumble strips 0-4 1 4+ 51 Edgeline 0-4 10 4+ 15 Non-edgeline 0-4 5 4+ 39 MN No rumble strips 0-4 17 Edgeline 4+ 1 4+ 2 Non-edgeline 0-4 3 4+ 28 MO No rumble strips 0-4 9 4+ 5 Edgeline 0-4 10 Non-edgeline 4+ 5 4+ 44 Rural two-lane roads PA No rumble strips 0-4 66 a Column indicates recovery area for treatment sites; shoulder width for nontreatment sites. Table 21. Distribution of combined rumble strip placement and recovery area or shoulder width by roadway type and state.

Crash type TOT FI SVROR SVROR FI Roadway type State Site type Treatment status Number of sites Length (mi) Mile- years Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) 27.013146.189272.113288.242520.281SRoNBA RS 53 35.51 129.58 394 3.04 162 1.25 198 1.53 97 0.75 No RS 37 19.88 194.52 1,325 6.81 648 3.33 556 2.86 265 1.36 Urban freeways PA CS RS 48 28.44 103.98 601 5.78 314 3.02 281 2.70 143 1.38 97.084279.171633.171432.4423,170.313SRoNBA RS 29 52.08 155.69 827 5.31 241 1.55 300 1.93 115 0.74 No RS 6 7.91 79.15 310 3.92 102 1.29 177 2.24 63 0.80 MO CS RS 12 25.65 77.90 200 2.57 64 0.82 107 1.37 42 0.54 56.02543.170157.00626.103190.08SRoNBA RS 18 15.43 57.02 107 1.88 44 0.77 63 1.10 28 0.49 No RS 16 25.23 245.62 429 1.75 207 0.84 273 1.11 136 0.55 Rural freeways PA CS RS 41 55.40 146.29 302 2.06 126 0.86 215 1.47 102 0.70 02.00233.03303.00327.02700.001SRoNBA RS 6 20.00 60.00 59 0.98 16 0.27 28 0.47 10 0.17 No RS 27 73.17 658.49 1,770 2.69 550 0.84 567 0.86 248 0.38 MN CS RS 27 103.32 508.56 1,205 2.37 373 0.73 476 0.94 193 0.38 46.00731.132199.080124.246220.901SRoNBA RS 14 19.11 51.18 196 3.83 66 1.29 114 2.23 44 0.86 No RS 12 15.04 150.37 458 3.05 122 0.81 152 1.01 65 0.43 MO CS RS 1 2.32 2.32 5 2.16 2 0.86 3 1.29 2 0.86 04.03170.15385.09144.17475.23SRoNBA RS 5 4.18 4.61 6 1.30 2 0.43 4 0.87 2 0.43 No RS 8 3.43 32.88 113 3.44 62 1.89 79 2.40 48 1.46 Rural multilane divided highways (nonfreeways) PA CS RS 4 1.18 2.35 6 2.55 2 0.85 3 1.28 1 0.43 50.04231.01612.000126.069214.874SRoNBA RS 28 95.51 285.67 220 0.77 70 0.25 43 0.15 15 0.05 No RS 28 86.66 776.29 515 0.66 199 0.26 162 0.21 76 0.10 MN CS RS 53 179.21 851.60 511 0.60 174 0.20 177 0.21 88 0.10 03.09146.01424.07202.17779.36SRoNBA RS 5 10.52 30.67 73 2.38 15 0.49 33 1.08 6 0.20 No RS 32 77.20 771.97 1,630 2.11 567 0.73 499 0.65 207 0.27 MO CS RS 1 2.33 2.33 2 0.86 0 0.00 0 0.00 0 0.00 74.04678.081147.010162.117181.631SRoNBA RS 20 23.34 69.90 86 1.23 56 0.80 41 0.59 24 0.34 Rural two-lane roads PA CS No RS 90 99.49 933.20 1,080 1.16 617 0.66 643 0.69 345 0.37 Table 22. Crash statistics by roadway type, state, site type, and treatment status.

Crash type TOT FI SVROR SVROR FI Roadway type State Site type Offset Number of sites Length (mi) Mile- years Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Non- edgeline edgeline edgeline edgeline edgeline edgeline edgeline 53 35.51 129.58 394 3.04 162 1.25 198 1.53 97 0.75 BA No RS 53 35.51 182.02 524 2.88 231 1.27 298 1.64 131 0.72 Non- 48 28.44 103.98 601 5.78 314 3.02 281 2.70 143 1.38 Urban freeways PA CS No RS 37 19.88 194.52 1,325 6.81 648 3.33 556 2.86 265 1.36 Edgeline 16 28.42 47.10 288 6.11 68 1.44 114 2.42 41 0.87 Non- 13 23.66 108.58 539 4.96 173 1.59 186 1.71 74 0.68 BA No RS 29 52.08 313.07 1,324 4.23 417 1.33 617 1.97 248 0.79 Edgeline 8 13.46 16.97 85 5.01 23 1.36 35 2.06 9 0.53 Non- 4 12.19 60.93 115 1.89 41 0.67 72 1.18 33 0.54 MO CS No RS 6 7.91 79.15 310 3.92 102 1.29 177 2.24 63 0.80 Edgeline 3 3.14 3.14 2 0.64 1 0.32 1 0.32 0 0.00 Non- edgeline 15 12.29 53.88 105 1.95 43 0.80 62 1.15 28 0.52 BA No RS 18 15.43 80.09 130 1.62 60 0.75 107 1.34 52 0.65 Edgeline 1 0.94 1.89 1 0.53 0 0.00 0 0.00 0 0.00 Non- 40 54.46 144.40 301 2.08 126 0.87 215 1.49 102 0.71 Rural freeways PA CS No RS 16 25.23 245.62 429 1.75 207 0.84 273 1.11 136 0.55 Edgeline 3 10.40 31.20 28 0.90 7 0.22 11 0.35 4 0.13 Non- edgeline 3 9.60 28.80 31 1.08 9 0.31 17 0.59 6 0.21 BA No RS 6 20.00 100.00 72 0.72 30 0.30 33 0.33 20 0.20 Edgeline 4 14.24 67.19 189 2.81 49 0.73 44 0.65 18 0.27 Non- 23 89.08 441.37 1,016 2.30 324 0.73 432 0.98 175 0.40 MN CS No RS 27 73.17 658.49 1,770 2.69 550 0.84 567 0.86 248 0.38 Edgeline 5 5.72 7.50 15 2.00 2 0.27 11 1.47 2 0.27 Non- 9 13.39 43.69 181 4.14 64 1.46 103 2.36 42 0.96 Rural multilane divided highways (nonfreeways) MO BA No RS 14 19.11 109.02 264 2.42 108 0.99 123 1.13 70 0.64 Non- edgeline 1 2.32 2.32 5 2.16 2 0.86 3 1.29 2 0.86 CS No RS 12 15.04 150.37 458 3.05 122 0.81 152 1.01 65 0.43 Table 23. Crash statistics by roadway type, state, site type, and edgeline vs. non-edgeline.

Edgeline 1 1.16 1.16 3 2.59 1 0.86 1 0.86 1 0.86 Non- edgeline edgeline edgeline edgeline edgeline 4 3.02 3.45 3 0.87 1 0.29 3 0.87 1 0.29 BA No RS 5 4.18 32.57 47 1.44 19 0.58 35 1.07 13 0.40 Non- 4 1.18 2.35 6 2.55 2 0.85 3 1.28 1 0.43 PA CS No RS 8 3.43 32.88 113 3.44 62 1.89 79 2.40 48 1.46 Edgeline 19 62.33 174.00 116 0.67 38 0.22 19 0.11 6 0.03 Non- 9 33.18 111.68 104 0.93 32 0.29 24 0.21 9 0.08 BA No RS 28 95.51 478.41 296 0.62 100 0.21 61 0.13 24 0.05 Edgeline 42 140.01 655.58 413 0.63 141 0.22 154 0.23 74 0.11 Non- edgeline 11 39.20 196.02 98 0.50 33 0.17 23 0.12 14 0.07 MN CS No RS 28 86.66 776.29 515 0.66 199 0.26 162 0.21 76 0.10 Edgeline 1 3.80 3.80 3 0.79 0 0.00 1 0.26 0 0.00 Non- 4 6.72 26.88 70 2.60 15 0.56 32 1.19 6 0.22 BA No RS 5 10.52 63.97 77 1.20 27 0.42 41 0.64 19 0.30 Non- edgeline 1 2.33 2.33 2 0.86 0 0.00 0 0.00 0 0.00 MO CS No RS 32 77.20 771.97 1,630 2.11 567 0.73 499 0.65 207 0.27 Edgeline 15 18.38 59.97 79 1.32 54 0.90 39 0.65 24 0.40 Non- 5 4.96 9.92 7 0.71 2 0.20 2 0.20 0 0.00 BA No RS 20 23.34 136.18 171 1.26 101 0.74 118 0.87 64 0.47 Rural two-lane roads PA CS No RS 90 99.49 933.20 1,080 1.16 617 0.66 643 0.69 345 0.37

Crash type Total (SVROR) Heavy vehicle Adverse pavement Low light Roadway type State Site type Treatment status Number of sites Length (mi) Mile- years Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) Total number of crashes Crash frequency (crashes/ mi/yr) 07.082127.013130.0646.189220.281SRoNBA RS 53 35.51 129.58 198 1.53 3 0.02 79 0.61 82 0.63 No RS 37 19.88 194.52 556 2.86 15 0.08 236 1.21 246 1.27 Urban freeways PA CS RS 48 28.44 103.98 281 2.70 42 0.40 111 1.07 133 1.28 57.053228.055274.064179.171670.313SRoNBA RS 29 52.08 155.69 300 1.93 51 0.33 108 0.69 103 0.66 No RS 6 7.91 79.15 177 2.24 26 0.33 76 0.96 82 1.04 MO CS RS 12 25.65 77.90 107 1.37 15 0.19 39 0.50 34 0.44 06.08416.09443.07243.170190.08SRoNBA RS 18 15.43 57.02 63 1.11 5 0.09 34 0.60 23 0.40 No RS 16 25.23 245.62 273 1.11 15 0.06 115 0.47 125 0.51 Rural freeways PA CS RS 41 55.40 146.29 215 1.47 15 0.10 125 0.85 90 0.62 31.03112.01250.0533.03300.001SRoNBA RS 6 20.00 60.00 28 0.47 1 0.02 24 0.40 5 0.08 No RS 27 73.17 658.49 567 0.86 14 0.02 308 0.47 266 0.40 MN CS RS 27 103.32 508.56 476 0.94 14 0.03 320 0.63 211 0.42 04.04454.09440.0431.132120.901SRoNBA RS 14 19.11 51.18 114 2.23 6 0.12 51 1.00 53 1.04 No RS 12 15.04 150.37 152 1.01 8 0.05 55 0.37 57 0.38 MO CS RS 1 2.32 2.32 3 1.29 1 0.43 2 0.86 1 0.43 34.04194.06130.0180.15375.23SRoNBA RS 5 4.18 4.61 4 0.87 1 0.22 2 0.43 1 0.22 No RS 8 3.43 32.88 79 2.40 12 0.37 40 1.22 33 1.00 Rural multilane divided highways (nonfreeways) PA CS RS 4 1.18 2.35 3 1.28 0 0.00 3 1.28 0 0.00 70.04360.07220.0731.01614.874SRoNBA RS 28 95.51 285.67 43 0.15 5 0.02 25 0.09 23 0.08 No RS 28 86.66 776.29 162 0.21 3 0.00 68 0.09 68 0.09 MN CS RS 53 179.21 851.60 177 0.21 10 0.01 119 0.14 77 0.09 41.0972.07160.0446.01479.36SRoNBA RS 5 10.52 30.67 33 1.08 3 0.10 19 0.62 8 0.26 No RS 32 77.20 771.97 499 0.65 26 0.03 160 0.21 192 0.25 MO CS RS 1 2.33 2.33 0 0.00 0 0.00 0 0.00 0 0.00 44.00653.08430.0478.081181.631SRoNBA RS 20 23.34 69.90 41 0.59 2 0.03 19 0.27 22 0.32 Rural two-lane roads PA CS No RS 90 99.49 933.20 643 0.69 28 0.03 281 0.30 322 0.35 Table 24. Crash statistics for supplemental analyses of SVROR crashes by roadway type, state, site type, and treatment status.

rumble strips. A basic negative binomial model was used with PROC GENMOD of SAS to estimate the regression coeffi- cients. A forward selection procedure was used to determine which variables to include in the SPF for each roadway type and state. The available variables are: ADT, outside RHR, and for divided highways, inside RHR. The following steps were used for variable selection: • Step 1: Estimate the intercept only. • Step 2: Estimate the intercept and the coefficient for the nat- ural log of the ADT (lnADT). If the coefficient for lnADT is positive, then include it in the SPF and continue to Step 3. • Step 3: Estimate the intercept and coefficients for lnADT and outsideRHR(RHROut).If thecoefficientfor lnADT is positive and the coefficient for RHROut is positive and has a p-value less than 0.15, then include lnADT and RHROut in the SPF. If not, include lnADT only in the SPF. For divided roads, if RHROut is included in the model, then continue to Step 4. • Step 4: Estimate the intercept and coefficients for lnADT, RHROut, and inside RHR (RHRIn). If the coefficient for lnADT is positive, the coefficient for RHROut is positive and has a p-value less than 0.15, and the coefficient for RHRIn is positive and has a p-value less than 0.15, then include lnADT, RHROut, and RHRIn in the SPF. If not, include lnADT and RHROut only in the SPF. In some cases, neither the coefficient of RHROut nor that of RHRIn met the above criteria; in those cases, the SPF reduces to the standard ADT-only model. In a few instances, the coef- ficient of ADT did not meet the above criteria. In those cases, an intercept-only or means model was selected for the SPF. The EB analysis was performed using the sum of the yearly crash frequencies at a given site during the before or after period. A factor was added to the model to account for the number of years at each site. Cross-Sectional Analysis Using Generalized Linear Model Analysis to Determine the Safety Effectiveness of Shoulder Rumble Strips on Different Roadway Types The evaluation of the safety effectiveness of shoulder rum- ble strips in the EB analysis discussed above is based solely on the comparison of sites with information from before and after the installation of rumble strips. Many additional sites are available that only have the information after the installation of the rumble strips (i.e., cross-sectional treatment sites). Since these sites provide additional information on the effectiveness of shoulder rumble strips, modeling efforts were undertaken to capitalize on all available information. A GLM with a negative binomial distribution and a log link was used to model the yearly crash counts. A repeated mea- sures correlation structure was included to account for the relationship in crashes at a given site across years (temporal correlation). A compound symmetry covariance structure was used. General estimating equations (GEE) were used to determine the final regression parameter estimates. The GEE regression model estimation technique has been demon- strated by Lord and Persaud (72). The selection of variables in the mean model was performed as described above with the addition of a factor for the presence of rumble strips (a 0,1 indicator variable) and a convergence criteria for the GEE. The GLM analysis was performed on two sets of sites: • All treatment and nontreatment sites (before-after and cross-sectional sites). • All before-after sites and all cross-sectional nontreatment sites (i.e., cross-sectional treatment sites are excluded). The first GLM analysis was performed to take advantage of all the data collected on the selected roadway types. The sec- ond GLM analysis was performed to provide a more direct comparison with the EB analysis results. In essence, the two methods use the same types of sites: the EB uses all nontreat- ment sites in the development of the SPFs and then the before and after sites in the safety evaluation; the second GLM analy- sis uses all nontreatment sites and the after sites more directly in the safety evaluation. A comparison of the three sets of results is discussed later in Analysis Results. Cross-Sectional Analysis to Determine the Impact of Rumble Strip Placement on the Safety Effectiveness of Shoulder Rumble Strips The analysis of the effect of rumble strip placement was performed using the same GLM approach used for the safety evaluation of shoulder rumble strips on different roadway types. All available site types were included in this analysis. The impact of offset distance was evaluated separately and then in combination with the recovery area. These analyses focused on SVROR FI crashes only. The rationale for this will become evident after reviewing the results of the safety eval- uations of shoulder rumble strips for TOT, FI, SVROR, and SVROR FI crashes for the four roadway types of interest. Both offset and recovery area were used as categorical variables as discussed earlier part in Descriptive Statistics. Three incrementally more detailed offset analyses were per- formed, based on various treatments of the offset and recovery area variables: • Comparison of edgeline and non-edgeline rumble strips against the no rumble strip category (i.e., no offset or non- treatment sites) (see Columns 3 and 4 in Table 20); 65

• Comparison of rumble strip offset at each of the three lev- els (0 to 8 in. [0 to 203 mm]; 9 to 20 in. [229 to 508 mm]; 21+ in. [533+ mm]) against the no rumble strip category (i.e., no offset or nontreatment sites) (see Columns 5 and 6 in Table 20); and • Comparison of the combination of rumble strip placement (i.e., edgeline; non-edgeline) and the recovery area at each of the five levels, including nontreatment sites with 4+ ft (1.2+ m) shoulders, against the nontreatment sites with 0 to 4 ft (0 to 1.2 m) shoulders (see Table 21). In all cases, a mean crash model was developed using the forward selection procedure discussed earlier and included ADT, RHROut, RHRIn, and the presence of rumble strips as long as their coefficients met the above criteria. Supplement Analysis for SVROR Crash Types Supplemental analyses of specific SVROR crash types (i.e., heavy vehicle, adverse pavement condition, and low-light condition) were also conducted using both the EB method- ology and the cross-sectional GLM methodology. For the EB methodology, separate SPFs were not developed for each crash type as described above. Instead, the SVROR SPF was used along with the percentage of crashes of the selected type to produce an SPF for each crash subtype. For crash type i, the SPF was defined as The percent (PCTi) is based on the accident count for site- years without rumble strips (i.e., the before period for before- after sites and cross-sectional sites without rumble strips) and is computed for each roadway type and state. The EB methodol- ogy is then carried out as described above. The cross-sectional GLM methodology was carried out in the same manner as for the roadway type evaluation, both with and without the cross- sectional rumble strip sites. The analysis of SVROR crashes involving heavy vehicles does not specifically account for heavy vehicle exposure. The ADT variable in the SPFs is for total traffic, including both pas- senger cars and heavy vehicles. Without specifically account- ing for heavy vehicle exposure, the baseline assumption of the analysis is that the percentage of heavy vehicle traffic, relative to the total traffic volume, remained constant throughout the analysis period. Similarly, the analysis of SVROR crashes during adverse pavement conditions does not specifically account for the variability in weather from year to year. Without specifically accounting for variability in weather, the baseline assumption of the analysis is that traffic is exposed to the same weather conditions from year to year throughout the analysis period. SPF PCT SPFi i SVROR= × Analysis Results Analysis results are first presented for estimating the safety effectiveness of shoulder rumble strips on different roadway types, followed by the analysis results for estimating the impact that rumble strip placement has on the safety effectiveness of shoulder rumble strips. The results of the supplemental analy- ses focusing on heavy vehicle, adverse pavement condition, and low-lighting condition crashes are presented last. Estimating the Safety Effectiveness of Shoulder Rumble Strips on Different Roadway Types Before-After EB Analysis Results The EB analysis consisted of the following two steps: • Develop SPF models based on all nontreatment sites. • Using the SPFs, evaluate the safety effectiveness of shoul- der rumble strips using crash data from the before-after sites only. SPF results. The approach discussed in the previous sec- tion was implemented for each crash type, separately for each roadway type and state. SPFs were developed for each crash type in each category, using ADT and RHR as predictor vari- ables. An attempt was made to include both outside and inside RHR in SPFs for divided highways. For undivided highways (i.e., rural two-lane roads in this analysis), only the outside RHR was included, if statistically significant. All non- treatment sites (i.e., nontreatment cross-sectional sites and before sites) were used for SPF development. Tables E-1 through E-4 in Appendix E summarize the crash frequency models for TOT, FI, SVROR, and SVROR FI crashes, respec- tively. The statistics shown for each roadway type and state SPF include the following: • Number of nontreatment sites; • Intercept: estimate and standard error; • lnADT coefficient: estimate, standard error, and p-value (or significance level); for example, a p-value of 0.05 or less indicates that the coefficient is significantly different from 0 at the 0.05 significance (or 95 percent confidence) level; • Outside RHR: estimate, standard error, and p-value; • Inside RHR (divided highways only): estimate, standard error, and p-value; • Model dispersion parameter: estimate and standard error; and • Model R2LR value: the likelihood ratio R 2 LR, a measure of model fit between 0 and 1. The closer the value is to 1, the better the fit of the model is to the data. 66

Each SPF is represented by the following equations: Divided highways: Rural two-lane roads: where ADT = average daily traffic volume for both directions of travel combined (veh/day) RHROut = average roadside hazard rating for the outside (right) side of a divided highway RHRIn = average roadside hazard rating for the inside (median) side of a divided highway a,b,c,d = coefficients whose estimates are shown in Tables E-1 through E-4 in Appendix E ADT was included in all models, regardless of its signifi- cance, as long as its coefficient was positive. A decision was made to select an ADT model with a nonsignificant positive slope rather than a simple means model after comparing their predicted crash frequencies over the range of ADTs in a given category. Including the before sites of the before-after site pairs in the SPF modeling ensured that no extrapolation out- side an ADT range would occur since ADTs changed only slightly from year to year. In only 6 of the 36 SPFs was ADT not significant at the 0.15 level. Outside and inside RHRs were only included in the SPF model for divided highways if they were significant at the 0.15 significance level. The selection of these two explanatory vari- ables was such that inside RHR was only included if it was sig- nificant and if outside RHR was significant and both were in the expected direction. Outside RHR was generally significant for Pennsylvania urban freeways and Minnesota and Mis- souri rural two-lane roads. Inside RHR was significant for selected crash types on Pennsylvania urban freeways only. Including RHR in the SPF allowed for a more accurate pre- diction of crashes in the after period, had no shoulder rum- ble strips been installed, as compared to using an ADT model only. In those cases where RHR was significant, not including it in the model could potentially result in wrongly attributing a safety improvement effect to rumble strips. The analyses of SVROR crashes included SVROR crashes to the right and to the left of the road. No effort was made to distinguish crashes by side of the road; however, by including RHR for both the outside and inside shoulder of divided highways, the analyses tried to account for the differences between SVROR crashes to the right and left. Also, it should be noted that for the states that treated both sides of a divided Expected crashes mi yr exp a blnADT cRHROut= + +( ) (2) Expected crashes mi yr exp a blnADT cRHR dROut = + + + HRIn( ) ( )1 highway as separate sites (i.e., Missouri and Pennsylvania), the RHR variables in the models represent the values for a sin- gle side of the divided highway. When both sides of a divided highway were treated as a single site (i.e., Minnesota sites), the RHR variables in the model represent average values for both directions of travel. Similarly, the RHR variable in the model for rural two-lane roads represents the average RHR for both sides of the roadway. Thus, the analysis tried to account for SVROR right and SVROR left crashes, without necessarily distinguishing between the two crash types. Safety effectiveness results. For each crash type, roadway type, and state, the safety effectiveness of shoulder rumble strips was estimated in accordance with the approach discussed earlier. The final results are shown in Tables 25 through 28 for TOT, FI, SVROR, and SVROR FI crashes, respectively. For each crash type, 12 separate analyses were performed across the four roadway types of interest, based on data for individual states and/or combined data across states. The statistics shown for each crash type, roadway type, and state (single or com- bined) are: • Number of treatment sites; • Total site length; • Percent change due to shoulder rumble strips: estimate and standard error; • Test statistic; and • An indication of whether rumble strips had a significant effect on the crash type of intercept. Four relevant findings from the EB analyses are noteworthy: • Of the 12 analyses based on TOT crashes (Table 25), 7 yield significant results at the 90 or 95 percent confidence level. Six of the seven analyses that indicate significant changes in crashes due to shoulder rumble strips result in counter- intuitive results, suggesting an increase in TOT crashes when shoulder rumble strips are installed. • Of the 12 separate analyses based on FI crashes (Table 26), only the analysis of Pennsylvania urban freeway data yields statistically significant results, indicating a decrease in FI crashes when shoulder rumble strips are installed on urban freeways. • Five of the 12 analyses based on SVROR crashes (Table 27) indicate that shoulder rumble strips have a statistically sig- nificant effect on SVROR crash frequency. Three of the sta- tistically significant results indicate a decrease in SVROR crashes when shoulder rumble strips are installed, while two indicate an increase in SVROR crashes when shoulder rumble strips are installed. • Six of the 12 analyses based on SVROR FI (Table 28) crashes indicate a statistically significant decrease in SVROR FI 67

Percent change in crash frequency from before to after rumble strip installation (%) Roadway type State Number of sites Total length (mi) Estimatea SEb Test statisticc Significance Urban freeways PA 53 35.5 –1.38 5.72 0.24 Not significant at 90% CL Combined 47 67.5 7.02 3.91 1.80 Significant at 90% CL MO 29 52.1 7.89 4.13 1.91 Significant at 90% CL Rural freeways PA 18 15.4 0.33 11.80 0.03 Not significant at 90% CL Combined 25 43.3 18.09 7.80 2.32 Significant at 95% CL MN 6 20.0 10.22 14.68 0.70 Not significant at 90% CL MO 14 19.1 22.00 9.46 2.32 Significant at 95% CL Rural multilane divided highways (nonfreeways) PA 5 4.2 –13.29 35.64 0.37 Not significant at 90% CL Combined 53 129.4 5.93 5.74 1.03 Not significant at 90% CL MN 28 95.5 14.38 8.01 1.80 Significant at 90% CL MO 5 10.5 40.49 18.00 2.25 Significant at 95% CL Rural two–lane roads PA 20 23.3 –24.40 8.61 2.83 Significant at 95% CL a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b SE: standard error of estimate. c Test statistic = abs(Estimate/SE); not significant at 90% CL if < 1.7; significant at 90% CL if ≥ 1.7; significant at 95% CL if ≥ 2. Percent change in crash frequency from before to after rumble strip installation (%) Roadway type State Number of sites Total length (mi) Estimatea SEb Test statisticc Significance Urban freeways PA 53 35.5 –16.01 7.25 2.21 Significant at 95% CL Combined 47 67.5 –6.88 5.88 1.17 Not significant at 90% CL MO 29 52.1 –5.84 6.41 0.91 Not significant at 90% CL Rural freeways PA 18 15.4 –12.61 14.62 0.86 Not significant at 90% CL Combined 25 43.3 –10.16 10.22 0.99 Not significant at 90% CL MN 6 20.0 –22.21 19.63 1.13 Not significant at 90% CL MO 14 19.1 –5.25 12.31 0.43 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 5 4.2 –40.12 42.52 0.94 Not significant at 90% CL Combined 53 129.4 –7.99 8.04 0.99 Not significant at 90% CL MN 28 95.5 5.13 12.66 0.41 Not significant at 90% CL MO 5 10.5 –19.24 21.82 0.88 Not significant at 90% CL Rural two–lane roads PA 20 23.3 –17.97 11.59 1.55 Not significant at 90% CL a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b SE: standard error of estimate. c Test statistic = abs(Estimate/SE); not significant at 90% CL if < 1.7; significant at 90% CL if ≥ 1.7; significant at 95% CL if ≥ 2. Table 25. Safety effectiveness of shoulder rumble strips on TOT crashes using the EB method. Table 26. Safety effectiveness of shoulder rumble strips on FI crashes using the EB method.

Percent change in crash frequency from before to after rumble strip installation (%) Roadway type State Number of sites Total length (mi) Estimatea SEb Test statisticc Significance Urban freeways PA 53 35.5 –5.81 7.32 0.79 Not significant at 90% CL Combined 47 67.5 –9.68 5.21 1.86 Significant at 90% CL MO 29 52.1 –7.91 5.71 1.38 Not significant at 90% CL Rural freeways PA 18 15.4 –17.71 12.27 1.44 Not significant at 90% CL Combined 25 43.3 40.01 12.40 3.23 Significant at 95% CL MN 6 20.0 38.36 26.62 1.44 Not significant at 90% CL MO 14 19.1 44.78 14.79 3.03 Significant at 95% CL Rural multilane divided highways (nonfreeways) PA 5 4.2 –25.46 37.44 0.68 Not significant at 90% CL Combined 53 129.4 –16.17 8.07 2.01 Significant at 95% CL MN 28 95.5 10.72 17.07 0.63 Not significant at 90% CL MO 5 10.5 16.87 21.76 0.78 Not significant at 90% CL Rural two–lane roads PA 20 23.3 –43.59 9.13 4.77 Significant at 95% CL a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b SE: standard error of estimate. c Test statistic = abs(Estimate/SE); not significant at 90% CL if < 1.7; significant at 90% CL if ≥ 1.7; significant at 95% CL if ≥ 2. Percent change in crash frequency from before to after rumble strip installation (%) Roadway type State Number of sites Total length (mi) Estimatea SEb Test statisticc Significance Urban freeways PA 53 35.5 –7.43 9.93 0.75 Not significant at 90% CL Combined 47 67.5 –17.14 7.30 2.35 Significant at 95% CL MO 29 52.1 –15.64 8.22 1.90 Significant at 90% CL Rural freeways PA 18 15.4 –23.20 15.71 1.48 Not significant at 90% CL Combined 25 43.3 –2.64 13.51 0.20 Not significant at 90% CL MN 6 20.0 –10.29 28.63 0.36 Not significant at 90% CL MO 14 19.1 0.16 15.84 0.01 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 5 4.2 –19.86 56.95 0.35 Not significant at 90% CL Combined 53 129.4 –36.42 9.71 3.75 Significant at 95% CL MN 28 95.5 –32.41 17.61 1.84 Significant at 90% CL MO 5 10.5 –44.59 23.16 1.93 Significant at 90% CL Rural two–lane roads PA 20 23.3 –36.66 13.35 2.75 Significant at 95% CL a A negative percent change indicates a decrease in crash frequency while a positive effect indicates an increase in crash frequency. b SE: standard error of estimate. c Test statistic = abs(Estimate/SE); not significant at 90% CL if < 1.7; significant at 90% CL if ≥ 1.7; significant at 95% CL if ≥ 2. Table 27. Safety effectiveness of shoulder rumble strips on SVROR crashes using the EB method. Table 28. Safety effectiveness of shoulder rumble strips on SVROR FI crashes using the EB method.

crashes when shoulder rumble strips are installed. The sta- tistically significant results are obtained for rural freeways and rural two-lane roads. Cross-Sectional GLM Analysis Results The safety effectiveness of shoulder rumble strips was also evaluated using a repeated measures analysis of variance approach based on all treatment and nontreatment sites as discussed in the Analysis Approach part of this section. This approach takes advantage of crash information on all study sites of interest. For each crash type, roadway type, and state, a regression model was developed to estimate crash frequency as a func- tion of ADT, outside RHR (and inside RHR when applicable), and the presence of rumble strips. Similar to the development of SPFs, a stepwise selection procedure was implemented to assess the significance of these variables, as discussed in the Analysis Approach part of this section. The GLM regression results for each crash type, roadway type, and state are shown in Tables F-1 through F-4 in Appendix F. These tables present the estimates of the regres- sion coefficients and their precision (standard error) along with their significance level, and the dispersion parameter sta- tistics. The introduction to Appendix F provides details on how to read and use these tables. The safety effectiveness of shoulder rumble strips is evalu- ated in Tables 29 through 32 for the four crash types of inter- est. These tables are directly obtained from Tables F-1 through F-4 by calculating the percent change due to rumble strips from the rumble strip coefficient shown in Appendix F. For each crash type (i.e., TOT, FI, SVROR, and SVROR FI crashes), 12 separate analyses were performed across the roadway types of interest, based on data for individual states and/or combined data across states. The statistics shown in Tables 29 through 32 for TOT, FI, SVROR, and SVROR FI crashes, respectively, include: • Number of sites (all treatment and nontreatment sites); • Number of site-years; • Percent change due to shoulder rumble strip: estimate and lower and upper 95 percent confidence limits; • Associated Type 3 p-value; and • An indication of whether rumble strips had a significant effect on the crash type of interest. A negative percent change in crash frequency indicates that crash frequencies decreased due to the shoulder rumble strip treatment; conversely, a positive change indicates an increase in crash frequencies. The 95 percent confidence limits of the percent change provide an assessment of whether the change, positive or negative, is statistically significant at the 95 percent confidence level: if the interval contains zero, then the change is not statistically significant (i.e., not different from zero) at the 95 percent confidence level; if the interval does not con- tain zero, then the change is statistically significant (i.e., dif- ferent from zero) at the 95 percent confidence level. The Type 3 p-values in the last column of Tables 29 through 32 also provide an indication of whether rumble strips have a significant effect on crash frequencies. These p-values correspond to the score statistics produced in the Type 3 GEE analysis and are generally more conservative than the p-values associated with the computation of the 95 per- cent confidence limits, which are computed with the Wald statistic. Generally, these two p-values are in agreement with each other; however, when the two disagree, the Type 3 p- value should be the one on which to base conclusions (73,74). In most cases in Tables 29 through 32, the Type 3 p-value is no more than 0.10 when the p-value associated with the con- fidence limits is 0.05. Several relevant findings from the cross-sectional GLM analyses based on all treatment and nontreatment sites are as follows: • Of the 12 analyses of TOT crashes (Table 29), 3 yield sig- nificant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Each of the statistically signif- icant results is for rural multilane divided highways (non- freeways), and each result is counterintuitive, indicating an increase in TOT crashes when shoulder rumble strips are installed. • Of the 12 analyses of FI crashes (Table 30), 2 yield signifi- cant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Both results indicate a decrease in FI crashes when shoulder rumble strips are installed. • Of the 12 analyses of SVROR crashes (Table 31), 6 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Three of the analyses indicate a decrease in SVROR crashes when shoulder rumble strips are installed, while the other three analysis results are counterintuitive. All of the counterintuitive results are based on data for rural multilane divided highways (non- freeways). • Of the 12 analyses of SVROR FI crashes (Table 32), 3 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. All three results indicate a decrease in SVROR FI crashes when shoulder rumble strips are installed. The GLM method was then repeated using a smaller set of sites consisting of all nontreatment sites and cross-sectional nontreatment sites. Thus, the cross-sectional treatment sites were excluded from the previous analysis. This analysis approach most resembles that of the EB analysis in that the two approaches use crash and site information from the same types of sites. 70

Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Significance Urban freeways PA 138 999 –3.8 –13.9 7.5 0.51 Not significant at 90% CL Combined 122 776 1.0 –11.8 15.5 0.89 Not significant at 90% CL MO 47 351 8.3 –4.8 23.3 0.26 Not significant at 90% CL Rural freeways PA 75 425 7.9 –12.6 33.1 0.50 Not significant at 90% CL Combined 104 788 20.1 2.8 40.2 0.03 Significant at 95% CL MN 60 424 16.4 –0.2 35.8 0.07 Significant at 90% CL MO 27 239 27.8 2.9 58.7 0.07 Significant at 90% CL Rural multilane divided highways (nonfreeways) PA 17 125 –19.2 –49.6 29.4 0.43 Not significant at 90% CL Combined 257 2,124 –14.0 –30.9 7.1 0.14 Not significant at 90% CL MN 109 726 –3.7 –16.2 10.6 0.59 Not significant at 90% CL MO 38 366 –16.7 –85.2 369.5 0.71 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –24.4 –48.1 10.1 0.14 Not significant at 90% CL Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Significance Urban freeways PA 138 999 –9.2 –22.2 6.0 0.23 Not significant at 90% CL Combined 122 776 –7.5 –19.7 6.6 0.28 Not significant at 90% CL MO 47 351 –2.6 –17.6 15.2 0.77 Not significant at 90% CL Rural freeways PA 75 425 –11.7 –30.8 12.7 0.32 Not significant at 90% CL Combined 104 788 0.6 –18.4 24.0 0.96 Not significant at 90% CL MN 60 424 6.9 –11.0 28.5 0.47 Not significant at 90% CL MO 27 239 5.3 –29.5 57.4 0.79 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 17 125 –41.5 –65.4 –1.2 0.09 Significant at 90% CL Combined 257 2,124 –27.5 –42.4 –8.6 0.01 Significant at 95% CL MN 109 726 –12.7 –27.6 5.3 0.16 Not significant at 90% CL MO 38 366 –39.7 –86.1 162.4 0.32 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –16.4 –45.6 28.5 0.36 Not significant at 90% CL Table 29. Safety effectiveness of shoulder rumble strips on TOT crashes based on all site types using the GLM method. Table 30. Safety effectiveness of shoulder rumble strips on FI crashes based on all site types using the GLM method.

Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p- value Significance Urban freeways PA 138 999 –3.7 –17.4 12.2 0.63 Not significant at 90% CL Combined 122 776 –9.5 –21.6 4.5 0.19 Not significant at 90% CL MO 47 351 –6.5 –19.0 7.9 0.40 Not significant at 90% CL Rural freeways PA 75 425 –2.1 –24.8 27.4 0.88 Not significant at 90% CL Combined 104 788 41.4 12.0 78.4 0.00 Significant at 95% CL MN 60 424 38.5 9.8 74.6 0.01 Significant at 95% CL MO 27 239 69.6 23.5 132.9 0.01 Significant at 95% CL Rural multilane divided highways (nonfreeways) PA 17 125 –23.3 –44.9 6.8 0.10 Significant at 90% CL Combined 257 2,124 –29.4 –49.0 –2.1 0.03 Significant at 95% CL MN 109 726 19.3 –7.8 54.4 0.19 Not significant at 90% CL MO 38 366 –8.9 –80.5 325.7 0.83 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –45.0 –64.6 –14.5 0.03 Significant at 95% CL Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p- value Significance Urban freeways PA 138 999 1.7 –15.7 22.6 0.87 Not significant at 90% CL Combined 122 776 –13.8 –27.0 1.7 0.09 Significant at 90% CL MO 47 351 –12.4 –27.6 6.0 0.20 Not significant at 90% CL Rural freeways PA 75 425 –13.0 –36.4 19.0 0.40 Not significant at 90% CL Combined 104 788 4.7 –19.9 36.7 0.72 Not significant at 90% CL MN 60 424 12.2 –14.1 46.6 0.39 Not significant at 90% CL MO 27 239 18.5 –21.7 79.2 0.41 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 17 125 –32.9 –61.3 16.4 0.18 Not significant at 90% CL Combined 257 2,124 –37.3 –54.3 –13.9 0.01 Significant at 95% CL MN 109 726 3.6 –26.4 45.7 0.85 Not significant at 90% CL MO 38 366 –59.4 –97.3 510.3 0.25 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –37.4 –61.1 0.8 0.06 Significant at 90% CL Table 31. Safety effectiveness of shoulder rumble strips on SVROR crashes based on all site types using the GLM method. Table 32. Safety effectiveness of shoulder rumble strips on SVROR FI crashes based on all site types using the GLM method.

73 The GLM regression results of this cross-sectional analysis are presented in Tables F-5 through F-8 in Appendix F for the four types of crashes. The structure of these tables is identical to that of Tables F-1 through F-4. The safety effectiveness of shoulder rumble strips is evaluated in Tables 33 through 36 for the four crash types of interest. These tables are directly obtained from Tables F-5 through F-8 by calculating the per- cent change due to rumble strips from the rumble strip coef- ficient shown in Appendix F. The general discussion of Tables 29 through 32 also applies to Tables 33 through 36. Similar to the previous two types of analysis, for each crash type (i.e., TOT, FI, SVROR, and SVROR FI crashes), 12 analyses are performed across the four roadway types of interest, based on data for individual states and/or combined data across states. The following relevant findings from the cross-sectional GLM analyses based on all before-after sites and all nontreat- ment cross-sectional sites (thus excluding cross-sectional treatment sites) are noteworthy: • Of the 12 analyses based on TOT crashes (Table 33), 3 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Each of these results is counter- intuitive. • Of the 12 analyses based on FI crashes (Table 34), 2 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Both results indicate a decrease in FI crashes when shoulder rumble strips are installed. • Of the 12 analyses based on SVROR crashes (Table 35), 4 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. Three of these results are counterintuitive. • Of the 12 analyses based on SVROR FI crashes (Table 36), 3 yield significant results at the 90 or 95 percent confidence level based on the Type 3 p-value. All three results indicate a decrease in SVROR FI crashes when shoulder rumble strips are installed. Comparison of Results from the Different Analysis Approaches The previous discussion covers two statistical methods— EB and GLM—applied to three different sets of data to assess the safety effectiveness of shoulder rumble strips on different roadway types. A direct comparison of the three sets of results is presented in Tables 37 through 40 for the four crash types, respectively. These tables are simply a side-by-side compila- tion of Tables 25 through 36, highlighting the relevant statis- tics for each crash type, roadway type, and state combination. A row is shaded in gray whenever a percent change obtained from any of the three analyses shows a statistically significant rumble strip effect at the 95 or 90 percent confidence level. Table 37 compares the results from the three analysis approaches to estimate the safety effectiveness of shoulder rumble strips based on TOT crashes. Comparisons of these results yield the following findings: • Analyses for rural multilane divided highways (nonfree- ways) yield the most consistent results across the three analysis approaches. Analyses for rural multilane divided highways (nonfreeways) based on the combined data for all three states and based on Missouri data yield statistically significant results; however, the results are counterintuitive. • Of the 36 analyses performed on TOT crashes using the three analysis approaches, only one (the EB analysis of rural two-lane roads based on Pennsylvania data) yields statistically significant results that appear intuitive (i.e., indicate a decrease in TOT crashes when shoulder rumble strips are installed). • Of the 36 analyses performed on TOT crashes using the three analysis approaches, 12 yield statistically significant results that are counterintuitive. • The largest range between the statistically significant esti- mates for the percent change due to rumble strips across analysis approaches is from 18.1 to 28.0 percent for rural multilane divided highways (nonfreeways) based on com- bined data. All analysis approaches indicate a significant increase in TOT crashes when shoulder rumble strips are installed. Table 38 compares the results from the three analysis approaches to estimate the safety effectiveness of shoulder rumble strips based on FI crashes. Comparisons of these results yield the following findings: • Of the 36 analyses performed on FI crashes using the three analysis approaches, 5 yield statistically significant results at the 95 or 90 percent confidence level. All of these results indicate a significant decrease in FI crashes when shoulder rumble strips are installed. • The largest range between the statistically significant esti- mates for the percent change due to rumble strips across analysis approaches is from −16.0 to −20.4 percent for urban freeways based on Pennsylvania data, suggesting close agreement between analysis approaches when statis- tically significant results are obtained. Table 39 compares the results from the three analysis approaches to estimate the safety effectiveness of shoulder rumble strips based on SVROR crashes. Comparisons of these results yield the following findings: • Of the 36 analyses performed on SVROR crashes across the three analysis approaches, 15 yield statistically significant

Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Significance Urban freeways PA 90 825 –5.2 –15.7 6.6 0.39 Not significant at 90% CL Combined 69 636 6.8 –7.4 23.2 0.36 Not significant at 90% CL MO 35 321 11.4 –2.5 27.4 0.15 Not significant at 90% CL Rural freeways PA 34 315 6.4 –20.6 42.5 0.69 Not significant at 90% CL Combined 72 646 28.0 8.4 51.2 0.02 Significant at 95% CL MN 33 291 16.7 –7.1 46.5 0.31 Not significant at 90% CL MO 26 238 27.7 2.7 58.8 0.08 Significant at 90% CL Rural multilane divided highways (nonfreeways) PA 13 117 –25.7 –59.6 36.4 0.37 Not significant at 90% CL Combined 203 1,871 –5.5 –27.7 23.5 0.65 Not significant at 90% CL MN 56 474 17.9 –0.6 39.7 0.09 Significant at 90% CL MO 37 365 –15.3 –86.2 421.0 0.74 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –24.4 –48.1 10.1 0.14 Not significant at 90% CL Table 33. Safety effectiveness of shoulder rumble strips on TOT crashes based on before-after sites and nontreatment cross-sectional sites using the GLM method. Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Significance Urban freeways PA 90 825 –20.4 –34.2 –3.7 0.02 Significant at 95% CL Combined 69 636 –4.1 –19.2 13.8 0.63 Not significant at 90% CL MO 35 321 –1.3 –18.3 19.1 0.89 Not significant at 90% CL Rural freeways PA 34 315 –8.6 –32.2 23.2 0.56 Not significant at 90% CL Combined 72 646 5.1 –24.3 45.8 0.75 Not significant at 90% CL MN 33 291 –17.7 –42.1 17.0 0.33 Not significant at 90% CL MO 26 238 1.9 –32.3 53.2 0.93 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 13 117 –43.8 –68.0 –1.2 0.10 Significant at 90% CL Combined 203 1,871 –14.4 –34.1 11.2 0.23 Not significant at 90% CL MN 56 474 7.0 –20.0 43.1 0.66 Not significant at 90% CL MO 37 365 –35.4 –85.5 188.1 0.40 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –16.4 –45.6 28.5 0.36 Not significant at 90% CL Table 34. Safety effectiveness of shoulder rumble strips on FI crashes based on before-after sites and nontreatment cross-sectional sites using the GLM method.

Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Significance Urban freeways PA 90 825 –9.6 –26.6 11.2 0.33 Not significant at 90% CL Combined 69 636 –9.5 –22.7 5.9 0.23 Not significant at 90% CL MO 35 321 –5.6 –18.8 9.7 0.50 Not significant at 90% CL Rural freeways PA 34 315 –13.0 –39.5 24.9 0.45 Not significant at 90% CL Combined 72 646 66.4 27.1 118.0 0.00 Significant at 95% CL MN 33 291 34.4 11.1 62.6 0.06 Significant at 90% CL MO 26 238 66.8 21.2 129.4 0.01 Significant at 95% CL Rural multilane divided highways (nonfreeways) PA 13 117 –22.7 –45.1 8.9 0.12 Not significant at 90% CL Combined 203 1,871 –25.9 –52.0 14.5 0.12 Not significant at 90% CL MN 56 474 12.1 –21.1 59.3 0.54 Not significant at 90% CL MO 37 365 –5.8 –80.1 346.1 0.89 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –45.0 –64.6 –14.5 0.03 Significant at 95% CL Percent difference in crash frequency with rumble strips present Roadway type State Number of sites Number of site-years Estimate (%) Lower 95% CL Upper 95% CL Type 3 p- value Significance Urban freeways PA 90 825 –9.5 –29.4 16.0 0.42 Not significant at 90% CL Combined 69 636 –17.5 –32.2 0.3 0.06 Significant at 90% CL MO 35 321 –13.7 –30.5 7.1 0.22 Not significant at 90% CL Rural freeways PA 34 315 –22.0 –47.0 14.7 0.22 Not significant at 90% CL Combined 72 646 14.7 –22.7 70.4 0.46 Not significant at 90% CL MN 33 291 –14.8 –41.1 23.4 0.40 Not significant at 90% CL MO 26 238 11.6 –25.8 67.7 0.59 Not significant at 90% CL Rural multilane divided highways (nonfreeways) PA 13 117 –21.5 –47.4 17.1 0.32 Not significant at 90% CL Combined 203 1,871 –39.6 –59.5 –9.9 0.02 Significant at 95% CL MN 56 474 –22.6 –52.5 26.1 0.36 Not significant at 90% CL MO 37 365 –56.9 –97.4 617.0 0.29 Not significant at 90% CL Rural two-lane roads PA 110 1,032 –37.4 –61.1 0.8 0.06 Significant at 90% CL Table 35. Safety effectiveness of shoulder rumble strips on SVROR crashes based on before-after sites and nontreatment cross-sectional sites using the GLM method. Table 36. Safety effectiveness of shoulder rumble strips on SVROR FI crashes based on before-after sites and nontreatment cross-sectional sites using the GLM method.

EB results (Table 25) GLM results using BA and nontreatment CS sites (Table 33) GLM results using all sites (Table 29) Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimate SE Test statistic Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Urban freeways PA 53 –1.4 5.7 0.24 90 –5.2 –15.7 6.6 0.39 138 –3.8 –13.9 7.5 0.51 Combined 47 7.0 3.9 1.80 69 6.8 –7.4 23.2 0.36 122 1.0 –11.8 15.5 0.89 MO 29 7.9 4.1 1.91 35 11.4 –2.5 27.4 0.15 47 8.3 –4.8 23.3 0.26 Rural freeways PA 18 0.3 11.8 0.03 34 6.4 –20.6 42.5 0.69 75 7.9 –12.6 33.1 0.50 Combined 25 18.1 7.8 2.32 72 28.0 8.4 51.2 0.02 104 20.1 2.8 40.2 0.03 MN 6 10.2 14.7 0.70 33 16.7 –7.1 46.5 0.31 60 16.4 –0.2 35.8 0.07 MO 14 22.0 9.5 2.32 26 27.7 2.7 58.8 0.08 27 27.8 2.9 58.7 0.07 Rural multilane divided highways (nonfreeways) PA 5 –13.3 35.6 0.37 13 –25.7 –59.6 36.4 0.37 17 –19.2 –49.6 29.4 0.43 Combined 53 5.9 5.7 1.03 203 -5.5 –27.7 23.5 0.65 257 –14.0 –30.9 7.1 0.14 MN 28 14.4 8.0 1.80 56 17.9 –0.6 39.7 0.09 109 –3.7 –16.2 10.6 0.59 MO 5 40.5 18.0 2.25 37 –15.3 –86.2 421.0 0.74 38 –16.7 –85.2 369.5 0.71 Rural two-lane roads PA 20 –24.4 8.6 2.83 110 –24.4 –48.1 10.1 0.14 110 –24.4 –48.1 10.1 0.14 EB results (Table 26) GLM results using BA and nontreatment CS sites (Table 34) GLM results using all sites (Table 30) Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimate SE Test statistic Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Urban freeways PA 53 –16.0 7.2 2.21 90 –20.4 –34.2 –3.7 0.02 138 –9.2 –22.2 6.0 0.23 Combined 47 –6.9 5.9 1.17 69 –4.1 –19.2 13.8 0.63 122 –7.5 –19.7 6.6 0.28 MO 29 –5.8 6.4 0.91 35 –1.3 –18.3 19.1 0.89 47 –2.6 –17.6 15.2 0.77 Rural freeways PA 18 –12.6 14.6 0.86 34 –8.6 –32.2 23.2 0.56 75 –11.7 –30.8 12.7 0.32 Combined 25 –10.2 10.2 0.99 72 5.1 –24.3 45.8 0.75 104 0.6 –18.4 24.0 0.96 MN 6 –22.2 19.6 1.13 33 –17.7 –42.1 17.0 0.33 60 6.9 –11.0 28.5 0.47 MO 14 –5.2 12.3 0.43 26 1.9 –32.3 53.2 0.93 27 5.3 –29.5 57.4 0.79 Rural multilane divided highways (nonfreeways) PA 5 –40.1 42.5 0.94 13 –43.8 –68.0 –1.2 0.10 17 –41.5 –65.4 –1.2 0.09 Combined 53 –8.0 8.0 0.99 203 –14.4 –34.1 11.2 0.23 257 –27.5 –42.4 –8.6 0.01 MN 28 5.1 12.7 0.41 56 7.0 –20.0 43.1 0.66 109 –12.7 –27.6 5.3 0.16 MO 5 –19.2 21.8 0.88 37 –35.4 –85.5 188.1 0.40 38 –39.7 –86.1 162.4 0.32 Rural two-lane roads PA 20 –18.0 11.6 1.55 110 –16.4 –45.6 28.5 0.36 110 –16.4 –45.6 28.5 0.36 Table 37. Comparison of results from three approaches to estimate safety effectiveness of shoulder rumble strips on TOT crashes. Table 38. Comparison of results from three approaches to estimate safety effectiveness of shoulder rumble strips on FI crashes.

EB results (Table 27) GLM results using BA and nontreatment CS sites (Table 35) GLM results using all sites (Table 31) Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimate SE Test statistic Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Urban freeways PA 53 –5.8 7.3 0.79 90 –9.6 –26.6 11.2 0.33 138 –3.7 –17.4 12.2 0.63 Combined 47 –9.7 5.2 1.86 69 –9.5 –22.7 5.9 0.23 122 –9.5 –21.6 4.5 0.19 MO 29 –7.9 5.7 1.38 35 –5.6 –18.8 9.7 0.50 47 –6.5 –19.0 7.9 0.40 Rural freeways PA 18 –17.7 12.3 1.44 34 –13.0 –39.5 24.9 0.45 75 –2.1 –24.8 27.4 0.88 Combined 25 40.0 12.4 3.23 72 66.4 27.1 118.0 0.00 104 41.4 12.0 78.4 0.00 MN 6 38.4 26.6 1.44 33 34.4 11.1 62.6 0.06 60 38.5 9.8 74.6 0.01 MO 14 44.8 14.8 3.03 26 66.8 21.2 129.4 0.01 27 69.6 23.5 132.9 0.01 Rural multilane divided highways (nonfreeways) PA 5 –25.5 37.4 0.68 13 –22.7 –45.1 8.9 0.12 17 –23.3 –44.9 6.8 0.10 Combined 53 –16.2 8.1 2.01 203 –25.9 –52.0 14.5 0.12 257 –29.4 –49.0 –2.1 0.03 MN 28 10.7 17.1 0.63 56 12.1 –21.1 59.3 0.54 109 19.3 –7.8 54.4 0.19 MO 5 16.9 21.8 0.78 37 –5.8 –80.1 346.1 0.89 38 –8.9 –80.5 325.7 0.83 Rural two-lane roads PA 20 –43.6 9.1 4.77 110 –45.0 –64.6 –14.5 0.03 110 –45.0 –64.6 –14.5 0.03 EB results (Table 28) GLM results using BA and nontreatment CS sites (Table 36) GLM results using all sites (Table 32) Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimate SE Test statistic Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Number of sites Estimate (%) Lower 95% CL Upper 95% CL Type 3 p-value Urban freeways PA 53 –7.4 9.9 0.75 90 –9.5 –29.4 16.0 0.42 138 1.7 –15.7 22.6 0.87 Combined 47 –17.1 7.3 2.35 69 –17.5 –32.2 0.3 0.06 122 –13.8 –27.0 1.7 0.09 MO 29 –15.6 8.2 1.90 35 –13.7 –30.5 7.1 0.22 47 –12.4 –27.6 6.0 0.20 Rural freeways PA 18 –23.2 15.7 1.48 34 –22.0 –47.0 14.7 0.22 75 –13.0 –36.4 19.0 0.40 Combined 25 –2.6 13.5 0.20 72 14.7 –22.7 70.4 0.46 104 4.7 –19.9 36.7 0.72 MN 6 –10.3 28.6 0.36 33 –14.8 –41.1 23.4 0.40 60 12.2 –14.1 46.6 0.39 MO 14 0.2 15.8 0.01 26 11.6 –25.8 67.7 0.59 27 18.5 –21.7 79.2 0.41 Rural multilane divided highways (nonfreeways) PA 5 –19.9 56.9 0.35 13 –21.5 –47.4 17.1 0.32 17 –32.9 –61.3 16.4 0.18 Combined 53 –36.4 9.7 3.75 203 –39.6 –59.5 –9.9 0.02 257 –37.3 –54.3 –13.9 0.01 MN 28 –32.4 17.6 1.84 56 –22.6 –52.5 26.1 0.36 109 3.6 –26.4 45.7 0.85 MO 5 –44.6 23.2 1.93 37 –56.9 –97.4 617.0 0.29 38 –59.4 –97.3 510.3 0.25 Rural two-lane roads PA 20 –36.7 13.3 2.75 110 –37.4 –61.1 0.8 0.06 110 –37.4 –61.1 0.8 0.06 Table 39. Comparison of results from three approaches to estimate safety effectiveness of shoulder rumble strips on SVROR crashes. Table 40. Comparison of results from three approaches to estimate safety effectiveness of shoulder rumble strips on SVROR FI crashes.

results at the 95 or 90 percent confidence level. Seven of the statistically significant results indicate a significant decrease in SVROR crashes when shoulder rumble strips are installed, while eight indicate a significant increase in SVROR crashes when shoulder rumble strips are installed. All of the statistically significant results that are counter- intuitive are for rural multilane divided highways (non- freeways). • The EB analysis is the only analysis approach that yields a significant result for rural freeways based on the combined data, indicating a significant reduction in SVROR crashes at the 90 percent confidence level. • When comparing the estimated percent change due to rumble strips between the analyses which yielded statisti- cally significant results, in several cases the estimated per- cent changes are very close. For example, the EB analysis indicates a 43.6 percent reduction in SVROR crashes for rural two-lane roads based on Pennsylvania data, while both GLM analyses indicate a 45 percent reduction in crashes. In other instances, the estimates are far apart. For example, the EB analysis indicates a 44.8 percent increase in SVROR crashes for rural multilane divided highways (nonfreeways) based on Missouri data, while the GLM analysis using all sites indicates a 69.6 percent increase in SVROR crashes. • When more than one analysis approach yields a significant result for a given analysis, the results are always in the same direction, either indicating a significant decrease (or increase) in crashes due to shoulder rumble strips. Table 40 compares the results from the three analysis approaches to estimate the safety effectiveness of shoulder rumble strips based on SVROR FI crashes. Comparisons of these results yield the following findings: • Of the 36 analyses performed on SVROR FI crashes using the three analysis approaches, 12 yield statistically signifi- cant results at the 95 or 90 percent confidence level, and all of the statistically significant results indicate a significant decrease in SVROR FI crashes when shoulder rumble strips are installed. • The EB and the GLM analyses for both data sets yield statis- tically significant results for the same three roadway types (i.e., rural freeways—combined data; rural two-lane roads— combined data; and rural two-lane roads—Pennsylvania data). The confidence levels among the analysis approaches (i.e., the EB analyses and the two GLM analyses) are slightly different, but the estimated percent change due to rumble strips for a given analysis (i.e., rural freeways—combined data; rural two-lane roads—combined data; and rural two- lane roads—Pennsylvania data) is relatively consistent across the three analysis approaches. From the comparison of the results from the three analy- sis approaches to estimate the safety effectiveness of shoul- der rumble strips for the four crash types of interest (i.e., TOT, FI, SVROR, and SVROR FI crashes), it is recom- mended that the focus of the analyses be on SVROR and SVROR FI crash results. The rationale for this recommen- dation is as follows. First, the primary purpose of shoulder rumble strips is to reduce SVROR crashes. Second, as indi- cated in the earlier part on Descriptive Statistics, no strong argument can be made to support why shoulder rumble strips would affect crashes other than SVROR crashes. Third, many of the analysis results for TOT crashes are counterintuitive. Therefore, the analysis results for TOT and FI crashes should be viewed with caution. In summary, the conclusions of safety evaluations of shoulder rumble strips will be based on results for SVROR and SVROR FI crashes. The conclusions also will focus on results obtained from analyses of combined data sets, rather than results for an individual state. In general, the results from across the three analysis approaches are most consistent when based on the combined data. The conclusions of the safety evaluation of shoulder rum- ble strips will further focus on results obtained from the EB analyses. The EB method is typically the method of choice to evaluate the safety effectiveness of a treatment when data are available for sites before and after a treatment. The steps of the method are straightforward; the method accounts for regression to the mean effects, and most researchers in the field of safety analysis consider this the most appropriate evaluation method. It is also the preferred method described in the forthcoming HSM for conducting safety evaluations. The GLM with the GEE estimation technique is less well known to non-statisticians, is less straightforward in its implementation, and thus is used less frequently in the field of safety analysis. However, the GLM method has a number of advantages over the EB method in that it uses site-year data and yearly changes in ADT; allows for quantification of site variability across years; does not rely on strict before- after data with a separate group of reference sites; accounts for sample size when testing for statistical significance; and is based on statistical theory. The decision to draw conclu- sions of the safety evaluations of shoulder rumble strips based upon the EB analyses was made primarily because this analysis approach is the preferred method of the forthcom- ing HSM, but analyzing the data using both the EB and GLM methods (and the two data sets for the GLM method) proved valuable in showing that the analysis approach can signifi- cantly impact the results of an evaluation; it also helped to illustrate concerns about drawing conclusions based upon TOT and FI crashes. In summary, the average safety effects of installing milled shoulder rumble strips on the following roadway types and 78

their associated standard errors (SE) are estimated to be the following: Rural Freeways: • 10 percent reduction in SVROR crashes (SE = 5) and • 17 percent reduction in SVROR FI crashes (SE = 7). Rural Two-Lane Roads: • 16 percent reduction in SVROR crashes (SE = 8) and • 36 percent reduction in SVROR FI crashes (SE = 10). The EB and GLM methods applied to these two crash types and roadway types provide very similar results, with the largest difference in results being for SVROR crashes on rural two-lane roads. Even though the results of the EB analysis of SVROR crashes for rural multilane divided highways show a statistically significant result at the 95-percent confidence level, the result is counterintuitive. This appears to be an anomaly in the data for this roadway type. This anomaly can- not be fully explained at this time. In part, a number of fac- tors working in combination could provide this result, such as (a) the unreliability related to PDO crash records, (b) the design features of this roadway type (e.g., roadside features, speed limit), or (c) the driver population along this roadway type. Also the sample size for this roadway type is relatively small, making it difficult to control for confounding factors. Thus, because the result of the EB analysis of SVROR crashes on rural multilane divided highways does not pass a face validity test (i.e., the results are in the expected direction and the magnitude appears reasonable), the result is viewed as an anomaly in the data and is not presented as a credible result. Comparison of Results with Previous Results The remainder of this section compares the results of the safety evaluation of milled shoulder rumble strips performed during this research to results from several previous studies. Two types of roadway were considered: rural freeways and rural two-lane roads. • Rural Freeways. Results from Griffith (1), believed to be the most reliable and definitive previous safety evaluation of shoulder rumble strips for rural freeways as indicated in NCHRP Report 617 (28), are compared to the results from this research. The following present both the crash reduc- tion effectiveness estimates for shoulder rumble strips and their standard errors: – Results from Griffith (1):  21 percent reduction in SVROR crashes (SE = 10) and  7 percent reduction in SVROR FI crashes (SE = 15). – Results from this research  10 percent reduction in SVROR crashes (SE = 5) and  17 percent reduction in SVROR FI crashes (SE = 7). Griffith (1) suggests greater reductions in SVROR crashes compared to results from this research, while this research suggests greater reductions in SVROR FI crashes than Grif- fith (1). Three points of interest when comparing the results from this research and Griffith (1) pertain to the analysis approaches, the standard errors, and the rumble strip type. The results from this research are based on the EB analysis methodology (and a cross-sectional approach) to analyze the data, while Griffith (1) utilized a comparison group (C-G) before-after study with yoked comparison sites to analyze the data. Additionally, the standard errors reported in this research are smaller than the standard errors reported by Griffith (1) for both crash types, indicating greater reliability in the results. It should also be noted that although Griffith reported an expected reduction of 7 per- cent in SVROR FI crashes on rural freeways due to the installation of shoulder rumble strips, the standard error of the reduction is 15 percent, indicating that statistically the reduction is not significantly different from zero. Griffith (1) also reported results for all freeways (i.e., both rural and urban combined); since this type of analysis was not per- formed under this research, no comparison is made for this category. Finally, Griffith (1) analyzed sites with rolled rumble strips, while for this research, all treatment sites had milled rumble strips, so the rumble strip type differed between the two studies. • Rural Two-Lane Roads. Results from this research for rural two-lane roads are presented with the results reported by Patel et al. (2), the only previous safety evaluation of shoul- der rumble strips on rural two-lane roads. Both analyses used the EB method to analyze the data. The following pres- ent both the crash reduction effectiveness estimates for shoulder rumble strips and their standard errors: – Results from Patel et al. (2):  13 percent reduction in SVROR crashes (SE = 8) and  18 percent reduction in SVROR FI crashes (SE = 12). – Results from this research:  16 percent reduction in SVROR crashes (SE = 8) and  36 percent reduction in SVROR FI crashes (SE = 10). The two safety evaluations report similar expected reduc- tions in SVROR due to shoulder rumble strips, while the results from this research indicate more than double the safety effectiveness of shoulder rumble strips in reducing SVROR FI crashes along rural two-lane roads compared to that reported by Patel et al. (2). The standard errors reported for both safety evaluations are comparable, but the expected values for SVROR FI crashes are considerably different. It should also be noted that Patel et al. (2) ana- lyzed only Minnesota data, while the analysis for this research was based on data from Minnesota, Missouri, and Pennsylvania. There does not appear to be any over- lapping of Minnesota sites being used in both studies, and 79

both studies are based upon sites with milled shoulder rumble strips. Considering that (a) NCHRP Report 617 (28) states that the Griffith (1) study is the most definitive study on the safety effects of shoulder rumble strips, (b) results from Griffith (1) are already incorporated into draft chapters of the forthcom- ing HSM, and (c) the likelihood that results from Patel et al. (2) could eventually be incorporated in the HSM, Table 41 presents updated AMFs for the safety effectiveness of shoul- der rumble strips. Results from this research are combined with results from Griffith (1) and Patel et al. (2) in an effort to provide reliable and comprehensive estimates on the safety effectiveness of shoulder rumble strips on rural freeways and rural two-lane highways. The results are combined in a man- ner consistent with the procedures for combining study results for incorporation in the HSM (65). Table 41 presents the safety effectiveness estimates of shoulder rumble strips in the form of AMFs for potential inclusion in future editions of the HSM. These estimates are viewed as the most compre- hensive and reliable estimates for the safety effectiveness of shoulder rumble strips to date. The safety effectiveness estimates for rural freeways in Table 41 are indicated as applying to shoulder rumble strips in general, rather than specifically to milled or rolled rum- ble strips. Milled rumble strips are currently used by most highway agencies and, therefore, most recent studies, including the current study, have focused on milled rum- ble strips. The previous research by Griffith (1), which has been viewed as the definitive research on this topic, addressed rolled rumble strips. While the alerting properties of milled and rolled rumble strips may differ to some extent, these differences would not be expected to have a major influ- ence on the safety effectiveness of shoulder rumble strips. Furthermore, when the Griffith results were combined with the results of this research, only minor differences in the safety effectiveness estimates were found. Therefore, it appears that the combined results from the Griffith study and the current research provide the best overall estimate for the safety effectiveness of shoulder rumble strips on rural freeways. Determine the Impact of Rumble Strip Placement on the Safety Effectiveness of Shoulder Rumble Strips The effect of shoulder rumble strip offset on SVROR FI crashes was evaluated based on all treatment and nontreat- ment sites using the approach discussed previously in the Analysis Approach part of this section. The analysis focused on SVROR FI crashes because this crash type and severity level yielded the most consistent results when analyzing the safety effectiveness of shoulder rumble strips on different roadway types. The results of the three types of offset analy- ses (i.e., cross-sectional GLM analysis) are presented next. Edgeline vs. non-edgeline rumble strip effect as com- pared to no rumble strips. The effect of rumble strip place- ment (i.e., edgeline defined as consisting of offset distances of 0 to 8 in. [0 to 203 mm] vs. non-edgeline defined as consist- ing of offset distances of 9 in. [229 mm] and greater) was eval- uated against the absence of rumble strips. This analysis was performed separately for each roadway type and state and all states combined. The GLM regression results of this cross-sectional analysis pertaining to ADT and outside RHR are presented in Table G-1; an introduction to Appendix G provides details on how to read and use this table. The structure of this table is iden- tical to that of Tables F-1 through F-8. The remainder of the regression model, that is, the rumble strip placement statis- tics, is presented in Table 42. For each combination of roadway type, state (combined or single), and rumble strip placement category, Table 42 shows the offset regression coefficient, its 95 percent confidence limits and p-value, and the Type 3 p- value. The discussion of significance provided for Tables 29 through 36 also applies to Table 42. For this offset analysis, a number of GLM models did not converge and therefore no reliable regression coefficients could be obtained. This is most often the case when only a few sites of a given type in a given state were available (refer to Tables 20 and 21). In Table 42, two sets of analyses are presented for rural two-lane roads. In the one analysis, all of the available data from Minnesota, Missouri, and Pennsylvania are included, resulting in a combined analysis based on 257 sites. The data 80 Treatment Roadway type Accident type and severity AMF SE SVROR 0.89 0.1 Shoulder rumble stripsa Rural freeways SVROR FI 0.84 0.1 SVROR 0.85 0.1 Shoulder rumble stripsb Rural two-lane roads SVROR FI 0.71 0.1 a AMF/SE based upon combined results for rolled shoulder rumble strips from Griffith (1) and for milled shoulder rumble strips from this research. b AMF/SE based upon combined results from Patel et al. (2) and this research. Table 41. AMFs for shoulder rumble strips recommended for inclusion in the HSM.

Percent difference in crash frequency with rumble strips present (%)a Roadway type State Number of sites Number of site–years Rumble strip placement Regression coefficient of rumble strip placement Estimate Lower 95% CL Upper 95% CL P– value Type 3 p–value Urban freeways PA 138 999 Non–edgeline 0.02 1.7 –15.7 22.6 0.86 0.87 Edgeline –0.34 –28.8 –51 3.5 0.08 Combined 122 776 Non–edgeline –0.09 –8.9 –23.7 8.6 0.30 0.07 b Edgeline –0.29 –24.8 –49.5 12 0.16 MO 47 351 Non–edgeline –0.03 –2.9 –17.5 14.3 0.73 0.22 Edgeline Rural freeways PAc 75 425 Non–edgeline Edgeline –0.29 –25.1 –46.5 4.9 0.09 Combined 104 788 Non–edgeline 0.10 10.3 –17.3 47.1 0.50 0.16 Edgeline –0.35 –29.6 –50.6 0.5 0.05 MN 60 424 Non–edgeline 0.21 23.4 –6.3 62.5 0.14 0.05 d Edgeline –0.54 –41.7 –63.5 –6.8 0.02 MO 27 239 Non–edgeline 0.25 28.3 –17.8 100.2 0.27 0.13 Edgeline 0.27 31 11.6 53.9 <.001 Rural multilane divided highways (nonfreeways) PA 17 125 Non–edgeline –0.65 –47.7 –73.5 3.2 0.06 0.09 b Edgeline –0.41 –33.3 –53.1 –5.2 0.02 Combined 257 2,124 Non–edgeline –0.63 –46.5 –67.1 –13.1 0.01 0.03 d Edgeline 0.08 8.3 –25.8 58 0.68 MN 109 726 Non–edgeline –0.09 –8.2 –43.9 50.4 0.74 0.82 Edgeline MOc 38 366 Non–edgeline Edgeline Rural two-lane roads PAc 110 1,032 Non–edgeline Edgeline –0.50 –39.2 –62.5 –1.5 0.04 Combined 204 1,872 Non–edgeline –0.54 –41.9 –69.4 10.0 0.10 0.05 d Edgeline –0.56 –42.9 –70.5 10.5 0.10 Rural two-lane roads e MN 56 474 Non–edgeline 0.10 10.5 –42.3 111.8 0.76 0.27 a Percent change is relative to no RS. b Significant at 90-percent confidence level. c GLM algorithm did not converge. d Significant at 95-percent confidence level. e Excludes 53 Minnesota nontreatment cross-sectional sites. Table 42. Safety effectiveness of rumble strip placement on SVROR FI crashes based on all sites using GLM method.

for rural two-lane roads were also analyzed without the 53 nontreatment cross-sectional Minnesota sites since these sites account for almost half of the Minnesota rural two-lane roads (see Table 14) and could unduly influence the results. This exclusion impacts only the Minnesota and the combined analysis results for rural two-lane roads, presented in the last two rows of Table 42. The most interesting results from this analysis are for rural freeways (combined) and for rural two-lane roads (combined). For rural freeways, the analysis of the combined data shows sta- tistically significant results at the 90 percent confidence level based upon the Type 3 p-value. The estimates for the percent change due to offset of a given distance are the following: • 28.8 percent reduction in SVROR FI crashes for edgeline rumble strips and • 8.9 percent reduction in SVROR FI crashes for non-edgeline rumble strips. For rural freeways, these results provide evidence that offset distance impacts the safety effectiveness of shoulder rumble strips. The results suggest that by alerting drivers sooner that their vehicles have left the travel lane, drivers are allotted more time to gain control of their vehicles and return safely to the roadway. As the rumble strips are placed farther from the edgeline, by the time drivers are alerted that their vehi- cles have left the travel lane they have less time to avoid hit- ting a roadside object, and therefore rumble strips placed further from the edgeline are less effective in reducing SVROR FI crashes. The analysis results of the combined data for rural two- lane roads indicate a slightly different safety effect of offset distance than for rural freeways. Focusing first on the analy- sis of combined data using all available sites (i.e., 257 com- bined sites), the results indicate a significant reduction in crashes for either edgeline or non-edgeline shoulder rumble strips. On average, a 33.3 percent reduction in SVROR FI crashes is found for edgeline shoulder rumble strips, while a 46.5 percent reduction in SVROR FI crashes is found for non- edgeline shoulder rumble strips. Given the relative magni- tude of the difference between these two estimates, there is not much difference as compared to the difference between the two estimates for rural freeways. The minimal difference between the two estimates for edgeline vs. non-edgeline shoulder rumble strips is clearer in the analysis based upon the 204 combined sites that excludes the 53 Minnesota non- treatment cross-sectional sites. The analysis results based upon 204 combined sites indicate a 39.2 percent reduction in SVROR FI crashes for edgeline rumble strips compared to a 41.9 percent reduction for non-edgeline rumble strips. Essen- tially, there is a 3 percent difference between the two esti- mates, but for practical matters this analysis indicates that offset distance does not impact the safety effectiveness of shoulder rumble strips along rural two-lane roads. Possible explanations for the difference in results between rural freeways and rural two-lane roads include the following: • Different driving populations; • Differences in driving behavior while driving along a multi- lane divided roadway where opposing traffic is separated by a median and adjacent traffic is traveling in the same direction versus driving along a two-lane road without any type of physical separation between opposing vehicles trav- eling in the opposite direction; • Differences in design standards (e.g., related to horizontal and vertical alignments); and • Extreme differences in the roadside characteristics that cannot be fully explained in the analyses. In an effort to account for differences between roadway types, an analysis of the effect of edgeline rumble strips ver- sus non-edgeline rumble strips, as compared to no rumble strips, was performed by combining the data for all the roadway types. State and roadway type differences were accounted for in the statistical model by fitting the inter- cept, lnADT coefficient, and outside RHR coefficient for each individual combination of state and roadway type. This was accomplished by including a number of inter- actions in the model in addition to the main effect, that is, rumble strip placement (i.e., edgeline, non-edgeline, and no rumble strips). The GLM regression results of this cross-sectional analysis, for all 621 sites combined, pertaining to ADT and outside RHR are presented in Table G-2. Since this cross-sectional analysis is based on a modification of previous models, a sep- arate introduction to Table G-2 provides details on how to read and use this table. The remainder of the regression model, that is, the rumble strip placement statistics, is pre- sented in Table 43. For each rumble strip placement, the table shows the regression coefficient on the natural log-scale and as percent change, its 95 percent confidence limits and p-value, and overall Type 3 p-value. Overall, the presence of rumble strips across all states and roadway types, while accounting for state, roadway type, ADT, and RHR differences, does not significantly impact SVROR FI crashes as indicated by the Type 3 p-value of 0.26. Table 43 does suggest the following: • Although only marginally statistically significant (i.e., p-value of 0.12), installing rumble strips within 0 to 8 in. (0 to 203 mm) of the edgeline reduces SVROR FI crashes by 14.4 percent as compared to not installing rumble strips. • On the other hand, placing rumble strips 9+ in. (229+ mm) or more from the edgeline has no overall effect on SVROR 82

FI crashes as compared to not installing rumble strips (p-value of 0.999.) • Combined, these two points could be interpreted to indi- cate that edgeline rumble strips are more effective than non-edgeline rumble strips because the edgeline rumble strips provide a reduction in SVROR FI crashes, while non- edgeline rumble strips have no effect (i.e., 0 percent reduc- tion). However, the overall results are not statistically significant (i.e., Type 3 p-value = 0.26), so this analysis does not provide definitive results regarding the impact that placement has on the safety effectiveness of shoulder rum- ble strips. Effect of offset distance at three levels as compared to no rumble strips. The effect of offset distance in three ranges was evaluated against the absence of rumble strips: • 0 to 8 in. (0 to 203 mm), • 9 to 20 in. (229 to 508 mm), and • 21+ in. (533+ mm). The GLM regression results of this cross-sectional analysis, pertaining to ADT and outside RHR, are presented in Table G-3. The structure of this table is identical to that of Table G-1 in Appendix G. The remainder of the regression model is pre- sented in Table 44. For each combination of roadway type, state (combined or single), and offset category, Table 44 shows the offset regression coefficient, its 95 percent confidence lim- its and p-value, and the Type 3 p-value. Table 44 presents the results of the analyses designed to determine the impact that rumble strip placement measured at three levels has on the safety effectiveness of shoulder rum- ble strips. The two most interesting results from this analysis are for rural freeways (combined data) and for rural two-lane roads (combined data). For rural freeways, the results from the combined data are not statistically significant at the 90 percent confidence level, but a Type 3 p-value of 0.14 indi- cates borderline significance (i.e., at the 85 percent confi- dence level). The estimates of the percent change due to offset of a given distance as compared to no rumble strips show the following: • 28.8 percent reduction in SVROR FI crashes when rum- ble strips are placed within 0 to 8 in. (0 to 203 mm) of the edgeline, • 10.4 percent reduction in SVROR FI crashes when rumble strips are placed within 9 to 20 in. (229 to 508 mm) of the edgeline, and • 7.4 percent reduction in SVROR FI crashes when rumble strips are placed 21+ in. (533+ mm) from the edgeline. Although these results are not statistically significant, the results are consistent with the analyses of edgeline vs. non- edgeline rumble strips for rural freeways shown in Table 42. The analysis results for rural two-lane roads (combined data) suggest, at first, the opposite effect from that for rural freeways. The estimates (in this case statistically significant at the 90 percent confidence level based on the Type 3 p-value) of the percent change due to offset of a given distance as com- pared to no rumble strips show the following: • 33.2 percent reduction in SVROR FI crashes when rum- ble strips are placed within 0 to 8 in. (0 to 203 mm) of the edgeline; • 37.7 percent reduction in SVROR FI crashes when rumble strips are placed within 9 to 20 in. (229 to 508 mm) of the edgeline; and • 56.7 percent reduction in SVROR FI crashes when rumble strips are placed 21+ in. (533+ mm) from the edgeline. These results suggest that installing the rumble strips further away from the edgeline improves the safety effectiveness of the rumble strips, which is counterintuitive to some extent. Upon further review, the p-value associated with the offset range of 21+ in. (533+ mm) is 0.12, indicating a marginally significant result. This result is likely due to sample size issues, in which case it makes sense to combine the two categorical offset ranges of 9 to 20 in. (229 to 508 mm) and 83 Percent difference in crash frequency with rumble strips present (%)a Number of sites Number of site-years Rumble strip placement Regression coefficient of rumble strip placement Estimate Lower 95% CL Upper 95% CL P-value Type 3 p-value Edgeline –0.156 –14.4 –29.6 3.9 0.12b621 4,687 Non-edgeline 0.0001 0.0 –10.3 11.5 0.999 0.26 a Percent change is relative to no RS. b Significant at 85 percent confidence level. Table 43. Overall safety effectiveness of rumble strip placement on SVROR FI crashes based on all sites combined using the GLM method.

Percent difference in crash frequency with offset of given distance (%)a Roadway type State Number of sites Number of site-years Offset (in.) Offset regression coefficient Estimate Lower 95% CL Upper 95% CL P-value Type 3 p-value 78.068.06.227.51–7.120.002-9999831APsyaweerfnabrU 0-8 –0.34 –28.8 –51.0 3.6 0.08d 9-20 –0.11 –10.4 –32.7 19.2 0.45 Combined 122 776 21+ –0.08 –7.4 –21.4 9.0 0.35 0.14 0-8 –0.28 –24.6 –49.5 12.6 0.17 9-20 –0.72 –51.1 –56.2 –45.4 <.001c15374OM 21+ 0.02 2.0 –13.1 19.8 0.81 0.13 0-8 9-20 Rural freeways PAb 52457 21+ 0-8 –0.30 –25.8 –46.9 3.8 0.08d 9-20 0.06 6.4 –18.0 38.1 0.64 Combined 104 788 21+ 0.15 15.8 –29.5 90.4 0.56 0.31 0-8 –0.35 –29.2 –50.2 0.4 0.05c 9-20 0.26 29.9 –1.0 70.6 0.06d42406NM 21+ –0.30 –25.9 –47.8 5.2 0.09d 0.02c 0-8 –0.54 –41.7 –63.5 –6.8 0.02c93272OM 21+ 0.25 28.3 –17.8 100.2 0.27 0.13 0-8 9-20 Rural multilane divided highways (nonfreeways) PAb 52171 21+ 0-8 –0.40 –33.2 –53.1 –4.9 0.03c 9-20 –0.47 –37.7 –60.4 –2.0 0.04cCombined 257 2,124 21+ –0.84 –56.7 –84.8 23.4 0.12 0.07d 0-8 0.10 10.0 –24.9 61.2 0.63 9-20 –0.24 –21.2 –55.3 39.0 0.41 627901NM 21+ 0.45 57.5 –26.0 234.9 0.24 0.54 0-8 MOb 66383 21+ 0-8 Rural two-lane roads PAb 230,1011 9-20 a Percent change is relative to no rumble strip. b GLM algorithm did not converge. c Significant at 95 percent confidence level. d Significant at 90 percent confidence level. Table 44. Safety effectiveness of rumble strip offset on SVROR FI crashes based on all sites using the GLM method.

21+ in. (533+ mm) into a single categorical level, as pre- sented in Table 42. Effect of combination of offset and recovery area as com- pared to nontreatment sites with narrow shoulders. The analyses described above consider the safety effect of rumble strip offset without considering the potential impact of the width of the recovery area. Therefore, the data were further ana- lyzed to estimate the combined effect of rumble strip offset and recovery area on SVROR FI crashes. The effect of five combina- tions of offset distance and recovery area (or shoulder width for nontreatment sites) on SVROR FI crashes was evaluated against the absence of rumble strips (category No. 6 below). The six combinations of rumble strip offset and recovery areas evalu- ated are as follows: No. RS Offset, in. (mm) Recovery area, ft (m) 1. Edgeline 0 to 8 (0 to 203) 4+ (1.2+) 2. Edgeline 0 to 8 (0 to 203) 0–4 (0–1.2) 3. Non-edgeline 9+ (229+) 4+ (1.2+) 4. Non-edgeline 9+ (229+) 0–4 (0–1.2) 5. No RS NA 4+ (1.2+) (shoulder width) 6. No RS NA 0–4 (0–1.2) (shoulder width) The GLM regression results of this cross-sectional analy- sis pertaining to ADT and outside RHR are presented in Table G-4 in Appendix G. The structure and discussion of this table is identical to that of Table G-1. The remainder of the regression model, that is, the statistics for the offset- recovery area combination, is presented in Table 45. For each combination of roadway type, state (combined or single), and offset and recovery area combination, Table 45 shows the regression coefficient, the estimated safety effect and its 95 per- cent confidence limits and p-value, and the Type 3 p-value. As before, a number of GLM models did not converge, and there- fore no reliable regression coefficients could be obtained for these cases. In general, no trends concerning the safety impacts of the interaction between offsets and recovery distances are observed from this analysis. In summary, comparing results from all the previous analy- ses (i.e., Tables 42 through 45) performed to assess the impact that placement has on the safety effectiveness of shoulder rumble strips, it appears that Table 42 provides the most reli- able results. The results of the analysis of edgeline rumble strips versus non-edgeline rumble strips as compared to no rumble strips reveals the following: • On rural freeways, edgeline rumble strips are more effective in reducing SVROR FI crashes than non-edgeline rumble strips, and • On rural two-lane roads, there is no difference in the safety effect of rumble strips placed close to the edgeline (i.e., edge- line rumble strips) compared to rumble strips placed further from the edgeline (i.e., non-edgeline rumble strips). The following is the rationale for determining that these results are the most definitive: • For rural freeways, all of the analyses were in agreement as to the direction of the effect that placement has on the safety effectiveness of shoulder rumble strips, whether the results are statistically significant at the 90 percent confi- dence level or higher, marginally significant, or not signif- icant. With the exception of one result based on a single state (MO in Table 44), all of the analyses (Tables 42 through 45) indicate that edgeline rumble strips are more effective in reducing SVROR FI crashes than non-edgeline rumble strips. Additionally, Table 42 provides statistically significant results (i.e., Type 3 p-value = 0.07), based upon the combined data. • The results for rural freeways are logical in that it makes sense that by alerting drivers sooner rather than later, they have more time to correct their steering and return to the travel lanes before encountering a roadside object. The high design policies for roadsides along rural freeways and high design policies for rural freeway alignments further support these results. • The primary issue to be addressed is determining whether rumble strips placed closer to the edgeline are more effec- tive in reducing SVROR FI crashes than shoulder rumble strips placed further away from the edgeline. The analysis of edgeline rumble strips versus non-edgeline rumble strips as compared to no rumble strips serves the purpose of answering this primary issue. The analyses on the “effect of offset distance at three levels as compared to no rumble strips” and the “effect of combinations of offset and recov- ery area as compared to nontreatment sites with narrow shoulders” are an attempt to further investigate and explain this issue; however, they are unnecessary in answering the primary issue at hand. The apparent conflict in the results for rural two-lane roads for Tables 42 and 44 is not an issue. Based upon the database available, the data do not support categorizing offset distance at three levels. Sample size issues limit the statistical validity of the results for Tables 44 and 45. Finally, the results for rural freeways and rural two-lane roads should not be viewed as being in conflict with one another. Rather, the results for rural freeways should be viewed as statistically significant results indicating that edgeline rumble strips are more effective in reducing SVROR FI crashes (i.e., a 28.8 percent reduction) than non-edgeline rumble strips (i.e., a 8.9 percent reduction), whereas for rural two-lane roads, the estimates of the safety effects of edgeline and non-edgeline rumble strips are so close (i.e., 39.2 percent reduction compared to a 41.9 percent reduc- tion) that, for all practical purposes, the placement of shoulder 85

Percent difference in crash frequency with combined effects of offset and recovery area (RA) (%)a Roadway type State Number of sites Number of site– years Offset x RA Combination Coefficient of offset by RA combination Effect (%) Lower 95% CL Upper 95% CL P– value Type 3 p–value Non–edgeline, 5 ft RA –0.38 –31.5 –54.0 2.1 0.06 Non–edgeline, 0–4 ft RA –0.42 –34.0 –53.0 –7.4 0.02 999831APsyaweerfnabrU No RS, 5+ ft shoulder –0.45 –36.2 –56.2 –6.9 0.02 0.36 Edgeline, 5+ ft RA Non–edgeline, 5 ft RA Non–edgeline, 0–4 ft RA Combined b 122 776 No RS, 5+ ft shoulder Edgeline, 5+ ft RA –0.29 –24.8 –49.5 12.0 0.16 MOc 15374 Non–edgeline, 5 ft RA –0.03 –2.9 –17.5 14.3 0.73 0.22 Edgeline, 5+ ft RA Non–edgeline, 5 ft RA Non–edgeline, 0–4 ft RA Rural freeways PAb 52457 No RS, 5+ ft shoulder Edgeline, 5+ ft RA –1.07 –65.7 –86.4 –13.7 0.02 Edgeline, 0–4 ft RA –0.04 –3.6 –29.3 31.6 0.82 Non–edgeline, 5 ft RA –0.65 –47.5 –78.2 26.0 0.15 Combined 104 788 No RS, 5+ ft shoulder –0.76 –53.1 –80.5 13.1 0.09 0.21 Edgeline, 5+ ft RA –0.35 –29.6 –50.6 0.5 0.05 MNc 60 424 Non–edgeline, 5 ft RA 0.21 23.4 –6.3 62.5 0.14 0.05 d Edgeline, 5+ ft RA –0.78 –54.2 –72.9 –22.6 0.00 Edgeline, 0–4 ft RA 0.10 10.4 0.9 20.8 0.03 Non–edgeline, 5 ft RA 0.34 40.7 1.9 94.5 0.04 93272OM No RS, 5+ ft shoulder 0.09 9.8 –22.1 54.9 0.59 0.33 Edgeline, 5+ ft RA –0.66 –48.2 –66.2 –20.7 0.00 Non–edgeline, 5 ft RA –1.53 –78.4 –88.5 –59.4 <.001 Rural multilane divided highways (nonfreeways) 52171AP No RS, 5+ ft shoulder –0.97 –62.0 –79.3 –30.3 0.00 0.14 Table 45. Safety effectiveness of combined rumble strip offset and recovery area on SVROR FI crashes based on all sites using the GLM method.

Edgeline, 5+ ft RA –0.57 –43.4 –65.1 –8.2 0.02 Rural two–lane roads Combined 257 2,124 Edgeline, 0–4 ft RA –0.63 –46.5 –75.4 16.3 0.12 0.10e Non–edgeline, 5 ft RA –0.78 –54.4 –72.8 –23.6 0.00 Non–edgeline, 0–4 ft RA –1.05 –65.2 –96.0 204.7 0.34 No RS, 5+ ft shoulder –0.32 –27.3 –50.8 7.4 0.11 Edgeline, 5+ ft RA 0.11 11.3 –51.7 156.6 0.80 Edgeline, 0–4 ft RA 0.24 27.1 –54.2 253.0 0.65 Non–edgeline, 5 ft RA –0.08 –7.8 –61.7 121.7 0.86 Non–edgeline, 0–4 ft RA 0.19 20.9 –61.4 279.1 0.75 627901NM No RS, 5+ ft shoulder 0.07 6.9 –54.9 153.4 0.88 0.98 Edgeline, 5+ ft RA Non–edgeline, 5 ft RA Non–edgeline, 0–4 ft RA MO b 38 366 No RS, 5+ ft shoulder Edgeline, 5+ ft RA Edgeline, 0–4 ft RA Non–edgeline, 5 ft RA PA b 110 1,032 No RS, 5+ ft shoulder RA = Recovery Area a Percent change is relative to no RS with 0–4 ft shoulder unless otherwise noted. b GLM algorithm did not converge. c Percent change is relative to no RS with 5+ ft shoulder. d Significant at 95 percent confidence level. e Significant at 90 percent confidence level.

rumble strips on rural two-lane roads has no impact on their safety effectiveness. Supplemental Analyses Additional analyses of SVROR crashes were performed to investigate the safety effect that shoulder rumble strips may have on the following: • Crashes involving heavy vehicles, • Crashes that occur under adverse pavement conditions, and • Crashes that occur during low-lighting conditions. Both the EB methodology and the cross-sectional GLM methodology were used in these supplemental analyses, as described in the Analysis Approach earlier in this section. Analyses results are provided for rural freeways and rural two-lane roads only. Heavy Vehicle Crashes Analysis results showing the safety effectiveness of shoulder rumble strips on SVROR crashes involving heavy vehicles are presented in Table 46. The results suggest that shoulder rum- ble strips installed on rural freeways reduce SVROR crashes involving heavy vehicles by 41 percent (based on the EB analy- sis). No significant effect is found for rural two-lane roads. As indicated earlier, the results of this analysis should be viewed cautiously because the analysis does not specifically account for heavy vehicle exposure, but rather assumes that the per- centage of heavy vehicle volume, relative to the total traffic volume, remained constant throughout the analysis period. Given the level of detail of this analysis, the major findings should be viewed in the following manner: • The primary issue to be resolved is whether shoulder rum- ble strips should be designed specifically taking into con- sideration heavy vehicles. One cannot state with certainty whether current rumble strip dimensions have been designed specifically based upon the needs of drivers of heavy vehicles. This is because it is still unclear, from a human factors perspective, what stimuli levels are neces- sary to alert drivers. In addition, the dynamic properties of passenger cars and trucks vary widely. What can be clearly stated, however, is that given the current dimensions of shoulder rumble strips installed along rural freeways that vary to some degree, the analysis results suggest a reduc- tion in SVROR crashes involving heavy vehicles due to the installation of shoulder rumble strips. This implies that the current dimensions of shoulder rumble strips installed along rural freeways provide sufficient levels of stimuli to alert inattentive and/or drowsy drivers of heavy vehicles and that it is not necessary to design rumble strip patterns that are “more aggressive” based strictly on the needs of drivers of heavy vehicles. • Concerning the accuracy of the estimated safety effect of shoulder rumble strips on SVROR crashes involving heavy vehicles (i.e., approximately a 40 percent reduction), it should be viewed with caution given the level of detail and assumptions of the analysis, but at this point it is the best estimate available for this crash type of interest. It should also be noted that all three analysis approaches provide very similar crash reduction estimates. • Concerning the safety effectiveness of shoulder rumble strips impacting SVROR crashes involving heavy vehicles on rural two-lane roads, it is not clear whether the non- statistically significant results are due to sample size issues, exposure, some combination of the two, or some other issues, but at this point there is no evidence that shoulder rumble strips installed along rural two-lane highways impact SVROR crashes involving heavy vehicles. Crashes Under Adverse Pavement Conditions Analysis results showing the safety effectiveness of the shoulder rumble strips on SVROR crashes occurring during adverse pavement conditions (i.e., wet, snow, ice) are pre- sented in Table 47. Statistically significant effects were found for both rural freeways and rural two-lane roads. In some cases, a significant reduction in SVROR crashes occurring under adverse pavement conditions was found, while in other cases a significant increase was found. Given that the analysis does not account for the potential differences in weather conditions from year to year, no definitive conclusions can be drawn from this analysis concerning the impact that shoulder rumble strips may have on SVROR crashes that occur under adverse pavement conditions. Crashes in Low-Lighting Conditions Analysis results showing the safety effectiveness of the shoulder rumble strips on SVROR crashes occurring during low-lighting conditions (i.e., dusk, dawn, dark) are presented in Table 48. Statistically significant effects were found for both rural freeways and rural two-lane roads. In all cases, a significant reduction in SVROR crashes occurring during low-light conditions was found. The EB results indicate that shoulder rumble strips installed on rural freeways reduce SVROR crashes that occur during low-lighting conditions by 27 percent. Similar estimates are provided for both GLM analyses. Considering that many inattention and drowsy driver crashes occur during late-night hours, these analysis results are confounded in that it cannot be distinguished 88

EB results GLM results using BA and nontreatment CS sites GLM results using all sites Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimatea (%) SE Test statistic Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Combinedb 47 –41.54 8.01 5.19 69 –41.6 –60.9 –12.9 0.01 122 –45.5 –61.9 –22.1 0.00 MOb 29 –41.66 8.38 4.97 35 –40.6 –56.4 –19.0 0.02 47 –41.3 –56.0 –21.6 0.01 Rural freeways PA 18 –40.81 26.97 1.51 34 –44.2 –94.1 426.5 0.53 75 –43.9 –84.5 103.0 0.36 Combinedc 53 82.64 58.41 1.41 203 257 MNc 28 181.99 126.29 1.44 56 109 MOc 5 138.52 142.01 0.98 37 38 Rural two–lane roads PA c 20 –19.12 57.52 0.33 110 110 a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b Shaded cells indicate a significant rumble strip effect at the 95 or 90 percent confidence level. c Empty cells indicate that the GLM algorithm did not converge. EB results GLM results using BA and nontreatment CS sites GLM results using all sites Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimatea (%) SE Test statistic Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Combinedb 47 –17.86 7.30 2.45 69 –10.9 –29.9 13.4 0.37 122 –5.9 –23.9 16.5 0.60 MOb 29 –23.13 7.75 2.99 35 –13.8 –32.3 9.6 0.30 47 –15.1 –32.8 7.4 0.23 Rural freeways PAb 18 4.30 20.02 0.21 34 3.2 –39.1 74.9 0.91 75 39.6 –1.8 98.3 0.10 Combined 53 17.90 15.38 1.16 203 5.2 –31.5 61.7 0.81 257 5.0 –20.7 39.0 0.73 MNb 28 50.15 30.23 1.66 56 37.9 –14.8 123.1 0.23 109 74.5 24.3 145.0 0.00 MOb 5 98.10 49.14 2.00 37 135.0 –43.9 885.1 0.27 38 118.0 –45.6 773.1 0.29 Rural two–lane roads PAb 20 –30.43 16.42 1.85 110 –27.8 –57.9 23.7 0.17 110 –27.8 –57.9 23.7 0.17 a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b Shaded cells indicate a significant rumble strip effect at the 95 or 90 percent confidence level. Table 46. Safety effectiveness of shoulder rumble strips on SVROR crashes involving heavy vehicles. Table 47. Safety effectiveness of shoulder rumble strips on SVROR crashes under adverse pavement conditions.

EB results GLM results using BA and nontreatment CS sites GLM results using all sites Percent change in crash frequency from before to after rumble strip installation (%) Percent difference in crash frequency with rumble strips present (%) Percent difference in crash frequency with rumble strips present (%) Roadway type State Number of sites Estimatea (%) SE Test statistic Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Number of sites Estimatea (%) Lower 95% CL Upper 95% CL Type 3 p–value Combinedb 47 –27.10 6.84 3.96 69 –23.3 –38.5 –4.5 0.02 122 –21.8 –35.1 –5.8 0.01 MOb 29 –25.86 7.64 3.38 35 –18.0 –36.3 5.6 0.15 47 –20.7 –36.0 –1.8 0.05 Rural freeways PAb 18 –32.62 15.16 2.15 34 –28.3 –52.1 7.4 0.11 75 –30.0 –52.7 3.5 0.07 Combinedb 53 –11.44 12.54 0.91 203 –24.3 –49.8 14.1 0.15 257 –30.1 –49.2 –3.8 0.03 MN 28 27.28 26.71 1.02 56 11.2 –30.1 77.0 0.67 109 11.2 –21.1 56.7 0.55 MO 5 19.71 44.60 0.44 37 54.7 –72.8 781.1 0.63 38 40.5 –75.0 688.6 0.70 Rural two–lane roads PAb 20 –37.54 13.71 2.74 110 –40.0 –64.9 2.5 0.05 110 –40.0 –64.9 2.5 0.05 a A negative percent change indicates a decrease in crash frequency while a positive percent change indicates an increase in crash frequency. b Shaded cells indicate a significant rumble strip effect at the 95 or 90 percent confidence level. Table 48. Safety effectiveness of shoulder rumble strips on SVROR crashes under low-lighting conditions.

between crashes where poor delineation may have been a pri- mary factor versus crashes where inattention and/or drowsi- ness may have been a primary factor. Also, no attempt was made in this analysis to account for adverse pavement condi- tions. Given the level of detail of this analysis and the con- founding issues, only a general conclusion can be drawn stating that shoulder rumble strips likely result in a positive safety benefit during low-lighting conditions by providing positive guidance along the travel lanes. Summary of Key Findings The primary objectives of the safety evaluation of shoulder rumble strips are to do the following: • Quantify the safety effectiveness of milled shoulder rumble strips on specific types of roads, and • Quantify the safety effectiveness of shoulder rumble strips placed in varying locations with respect to the edgeline. Supplemental analyses of SVROR crashes were performed to investigate the safety effect that shoulder rumble strips have on crashes that (a) involve heavy vehicles, (b) occur under adverse pavement conditions, and (c) occur during low-lighting conditions. Based upon the analysis results, the key findings from the safety evaluation of shoulder rumble strips are summarized as the following: • The most reliable and comprehensive estimates to date of the safety effectiveness of shoulder rumble strips on rural free- ways and rural two-lane highways are as follows: Rural Freeways: – Shoulder rumble strips [based on combined results from this research and Griffith (1)]:  11 percent reduction in SVROR crashes (SE = 6) and  16 percent reduction in SVROR FI crashes (SE = 8). Rural Two-Lane Roads: – Shoulder rumble strips [based on combined results from this research and Patel et al. (2)]:  15 percent reduction in SVROR crashes (SE = 7) and  29 percent reduction in SVROR FI crashes (SE = 9). • Analyses for urban freeways and rural multilane divided highways (nonfreeways) did not show statistically signifi- cant decreases in crash frequencies with the installation of shoulder rumble strips. • Limited mileage of shoulder rumble strips along urban multilane divided highways (nonfreeways), urban multilane undivided highways (nonfreeways), urban two-lane roads, and rural multilane undivided highways (nonfreeways) pro- hibited formal evaluation of the safety effectiveness of this treatment along these roadway types. • On rural freeways, rumble strips placed closer to the edge- line (i.e., edgeline rumble strips) are more effective in reducing SVROR FI crashes than rumble strips placed fur- ther from the edgeline (i.e., non-edgeline rumble strips). • On rural two-lane roads, there is no difference in the safety effect of rumble strips placed closer to the edgeline (i.e., edgeline rumble strips) as compared to rumble strips placed further from the edgeline (i.e., non-edgeline rum- ble strips). • On rural freeways, shoulder rumble strips resulted in an estimated reduction of SVROR crashes involving heavy vehicles of approximately 40 percent. • On rural two-lane roads, there is no evidence that suggests shoulder rumble strips may result in a reduction of SVROR crashes involving heavy vehicles. • Mixed results were observed from the analysis of SVROR crashes that occur under adverse pavement conditions, but it should be noted that the study did not attempt to account for the frequency of adverse pavement condition occurrences. • Shoulder rumble strips appear to provide a positive safety benefit during low-lighting conditions. 91