Investigating the Impact of Lack of Motorcycle Annual Average Daily Traffic Data in Crash Modeling and the Estimation of Crash Modification Factors

Chapter 4. Data Collection and Summary

This chapter describes the data collection procedures undertaken to support the analyses outlined in chapter 3. In general terms, these data are primarily roadway inventory, crash data, total traffic volumes, and motorcycle traffic volumes. Other data collection includes further information on motorcycle licenses, registrations, and other sociodemographic data at the county level.

To identify which States could provide a substantial sample of motorcycle volume counts, the project team asked various contacts for this information, primarily through the following channels:

• A request to the Transportation Research Board committee on motorcycles and mopeds, ANF30.
  
  
• Listserv communication with State highway safety engineers.
  
  
• Requests to members of the FHWA Development of CMFs Project Technical Advisory Committee.

These inquiries quickly confirmed that few States have substantial motorcycle volume estimates. For many States, the only available motorcycle counts come from permanent counting stations, where full class counts are performed and the interest is focused on truck traffic, not motorcycles. No attempt is made to estimate motorcycle counts for nearby segments in these instances.

There were, however, a number of States reported by these contacts as having a large number of classification count locations from permanent and short-term count programs. These States focused on for determining which States could provide the best datasets for the analyses.

The prioritization of States for data collection considered several factors, including the following:

• Availability of electronically linked and stored data on the roadway geometry, traffic volumes (including motorcycle volumes), and crash data for 5 years.
  
  
• For crash data, the necessary ability to identify motorcycle-involved crashes.
  
  
• For traffic data, classification counts, including motorcycles (a distinct class) and the total count for each site.
  
  
• Availability of data on variables useful for predicting motorcycle AADT, such as motorcycle licensure and socio-demographics.
  
  
• The ability to provide seasonal estimates of motorcycle volumes. These are known to fluctuate significantly, and it would be desirable to reflect this in the models developed.
  
  
• Geographic diversification by including States from northern and southern regions.

For each method applied, this section describes the source of data, variables acquired, and steps taken to assemble the data into an appropriate format for analysis.

Avenue A Databases

For developing the avenue A models, data were collected from Florida and Pennsylvania. The project team selected these States because they had a large number of locations with an estimated motorcycle AADT; were able to provide linkable roadway inventory, traffic, and crash data; and expressed an interest in providing data in a timely manner. They also provided a degree of geographical diversity. The project team acquired data from Virginia to validate the models developed.

The following sections provide further details on the data acquired and data manipulation.

Florida

Florida provided statewide roadway inventory, traffic volume, and crash data. The Florida Department of Transportation (FDOT) provided the roadway inventory and traffic volume data on the 2012 Florida Transportation Information DVD. Roadway and traffic volume data variables included the following:

• Segment location and length.
  
  
• AADT estimates for current and past years.
  
  
• Motorcycle AADT estimates for current and past years for those segments where an estimate exists.
  
  
• Number of lanes.
  
  
• Posted speed.
  
  
• Surface width.
  
  
• Functional class.
  
  
• Divided versus undivided.
  
  
• Inside and outside shoulder widths.
  
  
• Median width.
  
  
• Median type.
  
  
• Degree of curvature.

Crash data were provided for 2008 to 2012, including the following:

• Location.
• Date.
• Severity.
• First and second harmful events.
• Location type.
• Crash type.
• Vehicle types involved.
• Number of vehicles involved.

The query included total, motorcycle, multi-vehicle motorcycle, and single-vehicle motorcycle crashes.

Roadway identification numbers and mileposts linked all roadway, traffic, and crash data. Segments were defined so that all variables were homogeneous for the entire length.

Further data collected included the number of motorcycle licenses and registrations by county from 2008 to 2012. These data were available from the Florida Department of Highway Safety and Motor Vehicles and were linked to the other data by county.

Sociodemographic data were obtained from the U.S. Census Bureau at the county level. These variables included the following:

• Population by age.
• Population by sex.
• Percentage of population with high school or college degree.
• Mean travel time to work.
• Housing units.
• Home ownership rate.
• Persons per household.
• Median household income.
• Percentage of persons below poverty level.
• Non-employer establishments.
• Number of firms.
• Land area in square miles.
• Population per square mile.

These data were also linked to the other data by county. Segments were initially grouped into the following six categories for analysis:

• Rural freeway (type 1).
• Urban freeway (type 2).
• Rural arterial (type 3).
• Urban arterial (type 4).
• Rural collector/local (type 5).
• Urban collector/local (type 6).

Table 3 provides the variable definitions of the database created. Table 4 shows the total length of roadway and total number of crashes by crash type for the six site types in Florida. Table 5 through table 8 provide summary statistics for other variables included in the Florida dataset by site type.

1 mi = 1.6 km.

STD = Standard deviation.

a = PST045213.
b = AGE135213.
c = AGE295213.
d = AGE775213.
e = SEX255213.
f = EDU635213.
g = EDU685213.
h = LFE305213.
i = HSG010213.

j = HSG010213.
k = HSG445213.
l = HSD410213.
m = INC110213.
n = PVY020213.
o = NES010212.
p = SBO001207.
q = LND110210.
r = POP060210.

Pennsylvania

Pennsylvania provided statewide roadway inventory, traffic volume, and crash data. Roadway and traffic volume data variables included the following:

• Segment location and length.
• AADT estimates for current and past years.
• Motorcycle AADT estimates for current and past years for those segments where an estimate exists.
• Divided versus undivided.
• Surface width.
• Lane width.
• Posted speed.
• Number of lanes.
• Right shoulder type.
• Right shoulder width.
• Left shoulder type.
• Left shoulder width.

Crash data were provided for 2009–2013, including the following:

• Location.
• Date.
• Severity.
• Location type.
• Crash type.
• Vehicle types involved.
• Number of vehicles involved.

The query included total, motorcycle, multi-vehicle motorcycle, and single-vehicle motorcycle crashes. County, route number, segment number, and offset linked all roadway, traffic, and crash data. Segments were defined so that all variables were homogeneous for the entire length. The data for freeways are for one direction of travel only.

Further data collected included the number of motorcycle licenses and registrations by county from 2009–2013. These data were available from the Department of Motor Vehicles and linked to the other data by county.

The project team obtained sociodemographic data from the U.S. Census Bureau at the county level. These variables included the following:

• Population by age.
• Population by sex.
• Percent of population with high school or college degree.
• Mean travel time to work.
• Housing units.
• Home ownership rate.
• Persons per household.
• Median household income.
• Percent of persons below poverty level.
• Non-employer establishments.
• Number of firms.
• Lane area in mi².
• Population per mi².

These data were also linked to the other data by county. Segments were initially grouped into the following four categories for analysis:

• Rural freeway (type 1).
• Urban freeway (type 2).
• Rural non-freeway (type 3).
• Urban non-freeway (type 4).

Table 9 provides the variables’ definitions. Table 10 shows the total length of roadway and total number of crashes by crash type for the six site types in Pennsylvania. Table 11 to table 14 provide summary statistics for other variables included in the Pennsylvania dataset by site type.

Note: Type 1 and type 2 data are for one direction of travel only.
1 mi = 1.6 km.

A = PST045213.
B = AGE135213.
C = AGE295213.
D = AGE775213.
E = SEX255213.
F = EDU635213.
G = EDU685213.
H = LFE305213.

I = HSG010213.
J = HSG445213.
K = HSD410213.
L = INC110213
M = PVY020213.
N = NES010212.
O = SBO001207.
P = LND110210.
Q = POP060210.

Virginia

Virginia provided statewide roadway inventory, traffic volume, and crash data. The roadway inventory and total AADT data were provided from the Virginia Department of Transportation (VDOT) EYROAD data files. Motorcycle AADTs, where available, were queried specifically for the project and provided by VDOT staff. Crash data were downloaded from the VDOT Web site.

Roadway and traffic volume data variables included the following:

• Segment location and length.
• Functional class.
• AADT estimates for current and past years.
• Motorcycle AADT estimates for the most current year available.
• Surface width.
• Pavement width.
• Number of lanes.
• Shoulder width.
• Minimum and maximum median widths within the segment.

Crash data were provided for 2010–2014, including the following:

• Location.
• Date.
• Severity.
• Location type.
• Crash type.
• Vehicle types involved.
• Number of vehicles involved.

Total, motorcycle, multi-vehicle motorcycle, and single-vehicle motorcycle crashes were queried. All roadway, traffic, and crash data were linked together by a route name variable unique to each segment and milepost. Segments were defined so that all variables were homogeneous for the entire length.

Segments were grouped into the following four categories for validation:

• Rural freeway (type 1).
• Urban freeway (type 2).
• Rural arterial (type 3).
• Urban arterial (type 4).

Table 15 provides variable definitions in the database. Table 16 shows the total length of roadway and total number of crashes by crash type for the four site types in Virginia. Due to the low numbers of motorcycle crashes for rural and urban freeways (18 and 90, respectively), the data for freeways were not used for validation of the avenue A models. Table 17 and table 18

provide summary statistics for other variables included in the Virginia dataset for rural arterials (type 3) and urban arterials (type 4).

1 mi = 1.6 km.

Avenue B Databases

The databases used for the avenue B analyses are a combination of actual and simulated data. The roadway inventory used total AADT and motorcycle AADT collected for the avenue A methods in Florida and Pennsylvania. This ensured that realistic combinations of roadway geometry and traffic volumes were represented. For motorcycle crashes, the crash counts were simulated using the SPFs developed in the A1 models as a starting point. Further details on this approach follow for both the B1 and B2 models.

Model B1 Approach

The model B1 approach was to develop CMFs using the EB before-after method. Using the simulated data, a countermeasure was assumed with a known value of its CMF. The CMF was applied to the simulated after period crashes for a selection of sites. Then, the CMF was estimated twice, once using the motorcycle AADTs and once for total AADT. A comparison was then made to see how the lack of motorcycle AADT affected the estimate of the CMF and its variance. The following steps discuss this process in greater detail:

1. Estimate the existing roadway geometry and traffic volume file for a given class of road (e.g., rural freeway). The appropriate SPF from the A1 modeling, including the overdispersion parameter k, is used to estimate the mean motorcycle crash frequency for each site, m_u. This represents the expected mean value of all sites with the same road geometry and traffic volumes. The SPFs used are documented in chapter 5.
   
   
2. Estimate the site-specific mean crash frequency m_i by simulating a multiplier, r, to be applied to m_u. The multiplier r is generated from a gamma distribution with the shape and rate parameters equal to the overdispersion parameter, k. This is done using the equation: m_i = r × m_u.
   
   
3. Simulate the observed crash count at a site. The count in year j is assumed to follow a Poisson distribution about its mean, m_i. In this step, the number of motorcycle crashes for each site i in year j , X_ij, are simulated using the Poisson distribution and the site mean, m_i. Counts for 6 years are simulated.
   
   
4. Select a subset of the sites as the treatment group, with the remainder to be used as a reference group. For the treatment sites, a CMF is assumed for a fictitious treatment with a known value. The first 3 years are used as a before period and the second 3 years as an after period. The crash counts for the after period are simulated by first applying the CMF value to the site-specific mean, CMF x m_i. Then, simulate the after period crashes at the treated sites using this new expected mean as was done for all sites in step 4. For the treated sites, these new crash counts in the after period replace those simulated in step 4.
   
   
5. Use the reference group sites to estimate two SPFs to predict motorcycle crashes. The first uses motorcycle AADT as an exposure variable, and the second uses total AADT. The geometric and volume data adopted for the B1 analysis came from Florida. The A1 models developed for Florida did not always support both motorcycle and non-motorcycle volumes in the same model, and where they were both included, the model performance was roughly the same as the model including only motorcycle AADT. For this reason, the B1 analysis did not apply models including both motorcycle and non-motorcycle AADT.
   
   
6. Use the SPFs from step 6 with the treatment site data to estimate the expected number of crashes in the after period and estimate the CMF and its variance using the EB before-after study methodology.
   
   
7. Compare the estimated CMFs and standard errors (SEs) of these estimates from step 7 to each other as well as to the assumed CMF value to determine whether using total AADT SPFs is less accurate than using SPFs with motorcycle AADTs and, if so, the magnitude of this bias.
   
   
8. Repeat steps 2 through 8 multiple times so that conclusions can be made with confidence and have broad applicability.

Model B2 Approach

The model B2 approach was to develop CMFs using cross-sectional regression modeling. In this approach, an assumed relationship was defined between one or more geometric variables and added to the SPFs developed in model A1. The relationship was defined in terms of a CMF. This modified SPF was then used to simulate the motorcycle crash counts. Then, a GLM was used to re-estimate the SPF using motorcycle AADT and then using total AADT only. A comparison was then made to see how the lack of motorcycle AADT affected the estimate of the CMF and its variance. The following steps discuss this process in greater detail:

1. Adopt the existing roadway geometry and traffic volume file for a given class of road (e.g., rural freeway). Re-estimate the appropriate SPF from the A1 modeling, including the overdispersion parameter k, after first assuming a CMF and including this CMF in the SPF as an offset in the model formulation. This ensures that the predicted motorcycle frequencies are realistic, given the assumed CMF value.
   
   
2. Use the new SPF from step 1 to estimate the mean motorcycle crash frequency for each site, m_u. This represents the expected mean value of all sites with the same road geometry and traffic volumes.
   
   
3. Estimate the site-specific mean crash frequency m_i by simulating a multiplier, r, and applied to m_u. The multiplier r is generated from a gamma distribution with the shape and rate parameters equal to the overdispersion parameter, k. This is done using the equation: m_i = r × m_u.
   
   
4. Simulate the observed crash count at a site. The count is assumed to follow a Poisson distribution about its mean, m_i. In this step, use the Poisson distribution and site mean m_i to simulate the number of motorcycle crashes for each site for a 5-year period.
   
   
5. Use the data with simulated motorcycle crash counts to develop SPFs with the same model form as from step 1, first using motorcycle AADT as an exposure variable and then using total AADT.
   
   
6. Compare the CMFs and SEs of these estimates derived from the parameter estimates of the SPFs developed in step 6 to each other as well as to the assumed CMF value to determine whether using total AADT SPFs is less accurate than using SPFs with motorcycle AADTs and, if so, the magnitude of this bias. Repeat steps 3 through 7 so that conclusions can be made with confidence and have broad applicability.