Investigating the Impact of Lack of Motorcycle Annual Average Daily Traffic Data in Crash Modeling and the Estimation of Crash Modification Factors
Chapter 7. Limitations and Future Research Needs
This section summarizes existing data limitations and research gaps identified through the assessment of available data sources, analytical methods, and the evaluation results.
Data Limitations
With respect to developing SPFs and CMFs for motorcycle crashes, there are a number of data limitations related to traffic volumes (AADT), crash data, and roadway inventory data.
AADT Data
Technology
States collect traffic volume counts using a variety of techniques and technologies and convert these counts into estimated AADT volume estimates. With respect to motorcycle counts, NCHRP 08-36, Task 92, Counting Motorcycles, and NCHRP Report 760, Improving Motorcycle Travel Data – Data Collection Protocols and Field Tests, provide excellent detail on the strengths and weaknesses of the existing counter technologies.(29,30)
The second of these reports tested five detector technologies for their ability to accurately classify all vehicle types, specifically motorcycles. The five detector types tested were as follows:
- Infrared (IR) Classifier: This is a portable or fixed location system using IR beam interruption to identify and classify vehicles. The version tested in NCHRP 760 was The Infra-Red Traffic Logger (TIRTL).(30) The unit sends beams in four pathways (two perpendicular to the direction of travel and two at the diagonals). The receiver detects are records two timed events (beam interruption and re-establishment of the beam) as a vehicle passes through the four beam paths.
- Inductive loops/piezo electric sensors: The Traffic Detector Handbook: Third Edition describes the principal components of an inductive loop detector as loops of insulated wire placed in a slot sawed in the pavement and connected to an electronic controller unit.(31) As vehicles pass over the loop, they lower the inductance in the loop, and this change is recorded by the electronics in the controller unit. Each passage is time stamped. In units capable of classification counts, software is programmed to match the inductance changes over time with the pattern expected for each of the 23 FHWA-defined vehicle classification bins. Piezoelectric detectors are imbedded sensors that send an electronic pulse to the controller unit whenever an axle/tire travels over the sensor. The pulse varies by tire force, affecting the sensor. These sensors can be used in an array to provide vehicle classification data based on the weight and number of tires going over the sensor array.
- When combined, as is done in Virginia, the inductive loops and piezo sensors detect motorcycles with increased accuracy based on magnetic length (from the inductive loop) and rejection of energy detected from adjacent lanes (based on waveform analysis from the piezo).
- Magnetometers: Magnetometers work in a similar fashion to inductive loops in that they detect changes to a magnetic field as vehicles pass through the detection zone. Magnetometers installed in an array are used with accompanying software to classify vehicles by the timing and extent of changes in the magnetic field. Magnetometers are passive, meaning that a portion of the vehicle must pass over the detector. This makes this type of detector ideal (when installed in groups in a pattern that covers the lane correctly) for classification counts and for detecting vehicle spacing.
- Multi-technology system: This is a newer technology implemented specifically to count motorcycles. At the time of the research for NCHRP 760, the technology had not yet advanced to the point where it could reliably classify non-motorcycles.(30) As tested, it is a lower cost alternative than the other technologies.
- Tracking video: At the time of the research for NCHRP 760, video-based real-time counting and vehicle classification was undergoing significant changes.(30) The technology as tested used digital video processing algorithms to identify and classify vehicles based on shape profiles. Planned addition of IR video process would potentially improve the technology by allowing the unit to also detect the heat signatures of vehicles for use in classification.(30)
Table 85 compares the five technologies tested in NCHRP 760.(30) For each, there are two measures of accuracy: MC for motorcycle detection, and non-MC for detection of all other vehicles. Costs for two- and four-lane roadways are shown for initial installation. While there may be portable versions of many of the technologies, the only portable system tested was the TIRTL. The skill level required to set up and calibrate the detector and associated electronics/software is shown as well. Table 86 (table 2 from NCHRP 08-36) shows the technology used in 24 States.(29)
Table 85. Comparisons of five detector technologies from NCHRP 760.(30)
Technology |
MC Accuracy (Percent) |
Non-MC Accuracy (Percent) |
Initial Cost (Two-Lane) |
Initial Cost (Four-Lane) |
Portability |
Skill Level for Setupa |
IR Classifier |
95 |
98 |
$26,850 |
$26,850 |
Fixed/portableb |
Expert |
Inductive loops/ piezoelectric sensors (full lane-width) |
45 |
95 |
$33,000c |
$61,000 |
Fixed |
Field technician |
Magnetometers |
80 |
95 |
$10,204 |
$15,964 |
Fixedd |
Field technician |
Multi-technology system |
50 |
N/A |
$6,000 |
$12,000 |
Fixedd |
Field technician |
Tracking video system |
75 |
90 |
$15,000 |
$15,000 |
Fixedd |
Field technician |
aSetup skill level required: expert versus field technician with proper training.
bTIRTL is available as either portable or fixed, but only portable TIRTL was tested in this research.
cEstimated by Texas Department of Transportation: $61,000 total for four-lane site and $33,000 total for two-lane site.
dSome components could be portable, or the detector could be portable with modification.
N/A = Not available.
Table 86. Use of various detector technologies in 24 States as reported in NCHRP 08-36.(29)
Technology |
Short Counts |
Short Counts |
Continuous Counts |
Continuous Counts |
Tested |
Used |
Tested |
Used |
Intrusive |
Road tubes |
13 |
20 |
n/a |
n/a |
Piezoelectric cable |
3 |
4 |
9 |
17 |
Conventional inductive loops |
6 |
2 |
4 |
8 |
Piezoelectric film |
1 |
0 |
4 |
3 |
Inductive loop signatures |
1 |
0 |
2 |
1 |
Quadrupole loops |
1 |
0 |
1 |
0 |
Magnetometers |
1 |
0 |
2 |
0 |
Non-Intrusive |
Manual |
0 |
1 |
n/a |
n/a |
Radar |
7 |
3 |
4 |
5 |
Video |
1 |
2 |
2 |
1 |
IR, including TIRTLs |
5 |
0 |
4 |
3 |
Acoustic |
1 |
0 |
2 |
0 |
N/A = Not applicable.
As can be seen from table 86, road tubes are the most common type of detector for short counts—NCHRP 760 did not test the road tubes because they were not considered to be capable of providing accurate motorcycle classification counts.(30) Piezoelectric and inductive loops together account for the vast majority of detector installations in the 24 States that responded to the survey. Of the non-intrusive methods, the TIRTL is second-most common (after radar), but the use of the non-intrusive technologies still lagged far behind the inductive loops and piezoelectric installations as of 2010 when the project was completed.
NCHRP 760 concluded that the TIRTL was the most accurate system for detecting both motorcycles and non-motorcycles.(30) It has some higher skill level requirements than the other detectors, and a relatively higher price compared to the next-most accurate system, magnetometers. The two technologies that performed least well in the test (multi-technology and tracking video) are undergoing rapid technology improvements and may become more useful in the future. If accuracy is improved, then their pricing and required skill level for installation are attractive.
AADT Data Limitations
FHWA’s Traffic Detector Handbook provides details on the limitations of motorcycle count data available.(31) In general, the report finds the relatively small amount of metal in many motorcycles combined with the fact that many motorcyclists ride near lane lines in order to give themselves more time to avoid cars moving into their lanes means that inductive loop detectors and half-lane axle sensors often undercount motorcycles. When motorcycles ride in closely spaced groups, the closely spaced axles and cycles often confuse available traffic monitoring equipment that has not been designed to identify the resulting pattern of closely spaced axles and vehicles.(31)
The following summarizes the limitations of motorcycle AADT estimates in the United States:
- Permanent counter installations: Most permanent count system installations are not optimized for counting motorcycles. This is important because permanent counters are the source of most classification counts taken by States. There are two main concerns with current installations: calibration and detector configuration. Calibration refers to the sensitivity settings of the detector and software to arrive at correct bin assignments for each vehicle detected. The optimal calibration is different for each detector technology and will not be repeated here. However, it is important to recognize that detectors do go out of calibration and must be checked and adjusted or replaced. The software interprets input signals from the detectors and performs the assignment of vehicle counts into bins based on the detected size of the vehicle. Again, the technical details are different for each detector technology and will not be repeated here. The software is proprietary firmware with some user control over the associations between detected vehicle size and bin assignments. This is critical because when detecting motorcycles, the range of vehicle sizes can overlap with small cars, and the detector/software arrangement can mis-assign motorcycles to other vehicle classes when the motorcycles travel in groups. Depending on how the detectors are placed in the roadway, motorcycles traveling near the edge of a lane may be missed entirely.
- Temporary count technology: Temporary classification count methods and technology exist that can capture motorcycles accurately. This technology is not widely in use throughout the United States at present. The more accurate temporary installations are, unfortunately, more complex than older, more familiar technology. As a result, there is concern that it may not be installed properly in all instances or that it is more susceptible to failure (at least with respect to accurate motorcycle counting) than the older technology. As adoption of the new technologies is slow, there is not enough practical field data available to judge the benefit/cost of the newer technology. For the purposes of this summary, however, it is important to note is that there are relatively few reliable motorcycle classification counts outside of permanent count locations.
- Detector calibration: Taking temporary counts to be a negligible contribution to the motorcycle count data at present, the assessment of how accurate motorcycle AADT is in the United States requires detailed knowledge of just the permanent count locations. Installations and tracking of sensor calibration vary from State to State. It is difficult, if not impossible, to summarize how accurate motorcycle counts are at a national level. At the State level, it is possible to assess how well the existing count infrastructure matches the recently developed advice on detector types and configuration, though that evaluation has not yet been done for every State. Calibration of detectors also varies by State. Virginia, for example, keeps records that help to identify failing detectors in advance of a complete data loss. Other States may have this information, but Virginia supplies it as part of their count database so users can decide for themselves whether to use the data from any particular permanent count location.
Some detector types (such as piezo) have a failure mode that affects small vehicle classifications earliest and most severely. When a detector starts to fail, motorcycle and small car counts are affected most. If a State is not proactively keeping the detectors calibrated and replacing those that are failing, analysts may have to selectively exclude some count sites for some time periods based on suspicions of unreliable data.
- Motorcycles traveling in groups: This issue is related to sensor configuration and calibration. When motorcycles travel in groups, they often ride in parallel or staggered formations at the outsides of a single lane. Detector systems that are not specifically configured (as discussed in NCHRP 08-81) to correctly recognize these groupings do a poor job classifying motorcycles into the proper bin.(30,32)
- Count locations. NCHRP 08-81 addresses another concern often expressed by motorcycling advocates—that the count locations are not selected in a way that optimizes their reliability for estimating motorcycle AADT.(30,32) In particular, the concern is that recreational riders go places where classification counts are rarely collected. The result is that State-level aggregated counts are lower than they should be because a large proportion of motorcycle trips take place off of the facilities with the best, most accurate counting technology; many of those trips are on facilities that have no classification counts collected. Estimated factors may be applied to arrive at an estimated motorcycle AADT for those locations; however, there is some concern that the factors themselves are not accurate for the facilities in question. Middleton et al. were able to show that crash locations are a reasonable surrogate for motorcycle traffic volume such that States could use that information when deciding where to place counters or to evaluate whether the current count locations are sufficient to capture reliable area- or State-level motorcycle counts.(33) In their study, the authors found that the placement of counters does vary considerably by State with the conclusion being that each State should examine the spatial correlation of permanent count locations and motorcycle crash locations when deciding if they need to make additions or changes to the count locations. The findings indicated that motorcycle crash locations are not too different from crashes involving other vehicle types and that, in general, permanent counters on higher functional class roadways are likely to be sufficient. At least on an area-wide basis, the locations with high crash counts correlated well with locations of high motorcycle volume counts and high total traffic volume counts.
- Weekday versus weekend counts. This is an issue that affects temporary count locations. Typically, these counts are scheduled as mid-week counts because they are aimed at obtaining typical travel volumes at a location. While motorcycle crash data would indicate that weekday travel is an important component of all motorcycle travel, it is unquestionably true that a significant (though unknown) proportion of all motorcycle trips take place on the weekends. Of purely recreational motorcycle trips, the weekend proportion is much higher. Estimating weekend motorcycle travel is thus complicated by the fact that weekend factors developed based on other available data (i.e., from permanent count locations) is inaccurate for recreational trips because it captures mainly the travel on major roadways. Motorcyclists do not completely avoid those major roadways, of course, but much of the recreational motorcycle travel is on weekends and off the major roadways.
For overall traffic safety analysis, these shortcomings indicate that motorcycle counts may be inaccurate, and the inaccuracies vary among the States in ways that are difficult to assess or make adjustments for an analysis. The data from any specific site (most likely a permanent count location in a State) will have unknown under-reporting problems depending on how closely the installation matches the ideal design.(30) With few exceptions, the health of the detector—its current ability to correctly detect motorcycles—is unknown. Even if a detector’s installation was well designed originally, its current status might not be available to researchers who might use flawed data without knowing that the detector was beginning to fail during the period in which their data was collected. Ultimately, all of the problems would appear to lead to under-reporting of motorcycle counts, and thus an analyst might feel secure in viewing the data as a minimum. However, even that assurance may be misplaced. Detailed knowledge of the detector system’s software and any user-defined settings is needed before analysts could be comfortable that they know what happens with mistaken detections of various types and how the software is set to record those counts in the various classification bins.
Crash Data Limitations
Crash data standards, completeness, and accuracy vary among States. The difficulties using crash data in safety analyses are well known but not particularly well documented in the literature. Practitioners are, however, well aware of the following data limitations:
- Accident reports and data definitions: Each State creates its own police accident report (PAR) and decides what data elements to include in its centralized State crash database. There is a national guideline—the Model Minimum Uniform Crash Criteria (MMUCC)—which provides a standardized set of data definitions for a minimum set of 110 data elements. However, this is a voluntary guideline, and most States’ PAR and crash database are below 100 percent compliance with MMUCC. Analysts have to understand the specific data element definitions of each State’s data that was used. Depending on which data elements are considered in the analysis, the ability to generalize the results from one State to another, or to the national picture, can be quite limited.
- Location accuracy: States generally do an excellent job of locating crashes on the roadway network for any State-maintained roads and for any roadways eligible for funding under the Highway Safety Improvement Program (HSIP). Recent FHWA requirements for each State to have an all-public-roads linear referencing system (LRS) and basic roadway inventory data means that soon all States will be able to place every crash in a single LRS. Today, however, there are still States that have not achieved that step of assigning LRS location codes to every crash. The result is that safety analysis over the entire network is complicated by the fact that only some crashes are readily associated with AADT and roadway descriptive data (for example, in a geographic information system (GIS) implementation of the State-roads LRS). Local roadway crashes may be located only using a non-linear location coding scheme that is incompatible with the State’s GIS and LRS. As a result, the ability to access and use local roadway crashes is impeded. This may not be a problem for some types of motorcycle safety analyses (e.g., ones focusing on motorcycle crashes on interstates), but it can be a serious barrier to analyzing motorcycle crashes in intersections at the local level or rural crashes on low-volume, low-functional class roadways. These crashes also tend to be the ones that State crash location staff spend the most time trying to correct. Local law enforcement agencies typically use local designations for roadways in their jurisdiction and may not provide the information that would allow the State-level staff to identify the location on the official State GIS or LRS. As discussed earlier, these are also the locations that are least likely to have a classification count available.
- Crash severity: Most (but not all) States use a variation on the KABCO injury severity scale to code personal injuries in crashes. K, A, B, and C are coded as injuries to individuals involved in the crash and correspond with the overall crash severity—the highest severity single injury in the crash is assigned as the overall crash severity. The definitions of these terms can vary markedly between States. Fatalities are typically defined based on the FARS criterion of a crash-related death within 30 days of the event. A-level injuries are the most severe level, but the descriptions may include “serious,” “incapacitating,” or other terms. B-level injuries are typically defined as apparent injuries at a lower or moderate level, with “non-incapacitating” and “apparent” often used as descriptors. The C-level injury definition may include descriptors such as “slight” or “possible.” PDO crashes are those in which there was no personal injury recorded but enough property damage to make the crash reportable under the State’s threshold criterion. Those criteria also vary among States and in interpretation among law enforcement agencies within a State.
Typically, States set a minimum dollar amount threshold plus any injury or fatality so that, in theory, all KABC crashes should be reported along with any crashes with property damage about the threshold dollar amount. In practice, the threshold dollar amount is treated as a rough approximation, and each agency’s practices dictate how its officers interpret the physical damage to vehicles and other property to arrive at the decision of whether or not to report a PDO crash. Another factor affecting reported motorcycle crashes is that many single-vehicle motorcycle crashes are unreported.
Safety analysis addresses the variation in severity codes in two basic ways. One way is to use all crashes. In practice, this equates to using all reported crashes, which in turn means that there is likely to be some systematic under-reporting. If safety comparisons are planned among jurisdictions, using all levels of crash severity can cause problems if the law enforcement agencies interpret the crash reporting threshold differently across those jurisdictions. For that reason, and because data quality is usually best for reports of serious crashes, analysts sometimes concentrate on “serious” crashes by taking fatalities and A-level injuries together or sometimes combining K, A, and B injury crashes. This has the advantage of providing greater comparability across jurisdictions but at the price of missing the majority of reported crashes. Typically, PDO crashes account for about 60 to 70 percent of all crashes. Analyses focused on crash locations and their attributes, excluding PDO and C-level injury crashes, are likely to rob the analysis of generalizability.
No description of crash severity would be complete without also pointing out that the KABCO values are determined by law enforcement officers, not trained medical personnel. When crash data and medical records are linked (as has been done many times in several States), the results point to a large discrepancy between the officers’ judgements of injury severity and the actual injury severity coded based on medical injury severity codes, medical treatments, or cost of treatment. As might be expected, fatalities are judged accurately most of the time (though there are cases of death away from the scene that are sometimes missed as well as successful patient resuscitations away from the scene). A, B, and C injury codes are often medically incorrect, so the calculated crash severity scores based on those injury codes are prone to error. These errors do not pose a large concern for most safety analyses because the errors are usually randomly distributed and the comparisons being made in the typical safety analysis would not be expected to change much. Unfortunately, typical motorcycle crash-related injuries are of a nature where the differences between officers’ and medical judgements are largest. If an analysis requires comparison of motorcycle crashes to all other vehicles’ crashes, the possibility exists that the KABC assignments to the motorcycle crashes are less accurate than the ones for occupants of other vehicles.
- Crash contributing factors, harmful events, and characteristics: Safety data analyses sometimes must obtain roadway and human factors information based solely on the PAR. For example, if the State lacks a robust roadway inventory system that can supply location descriptions, the only information about the crash site will come from the PAR.
The officers record those circumstances that they judge to have contributed to the crash, the apparent sequence of events, most harmful events, etc. When there are inconsistencies in data definitions, the ability to reliably aggregate data and form valid comparisons is limited. In addition, different States record different aspects of the crash, including the basic location descriptions. Most States collect a minimum set of roadway attributes, but those attributes differ widely among the States. If the analysis relies on knowing something about the roadway attributes, or circumstances of the crash, the analyst must know details about the data definitions as set by the State and, preferably, have some measures of accuracy and completeness available in order to judge the sufficiency of the database to support the intended analysis.
Roadway Inventory Limitations
Safety analyses often make use of roadway inventory data in order to understand the roadway attributes’ association with crash risk. There are a few notes worthy of consideration because they relate to AADT and crash data and the statewide roadway inventory data.
- Roadway class/ownership: Because of the limitations of AADT data and in some States the limited ability to locate local crashes on a statewide LRS, safety analyses are sometimes limited to the State-maintained/HSIP-eligible portions of the roadway network. As noted earlier, this selective use of higher functional classification roadways can miss an important subset of motorcycle crashes. Crashes on low-volume rural roads or at local roadway locations are more likely to be unlinked to any available roadway characteristics data. If the analysis focuses on specific roadway characteristics, that data may be missing for the local roadways.
- Missing data: As noted earlier, many State LRSs lack locations for local and low-volume roadways, although the systems are adding this information now. A related problem is that even when the LRS includes those locations, the roadway inventory file may have blanks for some of the key data. Just as AADT is less readily available for local and low-volume roadways in many States, so too are the roadway inventory data less complete. When the analysis requires knowledge of roadway attributes, the ability to compare jurisdictions may be impaired if too much of the detailed inventory data is missing.
Analysis Limitations
As discussed in chapter 2, most of the current research concerning motorcycle crashes has focused on discrete outcomes (i.e., the probability of a given crash severity, presence of a roadway or traffic control feature given that a crash has occurred, or probability of injury severity given that feature or crash type). Very little research focuses on developing SPFs or CMFs specifically for motorcycle crashes.
The same analysis methods available for estimating SPFs and CMFs for other crash types are applicable for motorcycle crashes. While having estimates of motorcycle volumes is preferred, the results of the analyses undertaken for this project indicate that using total AADT volumes is a reasonable substitution when motorcycle AADT is not available.
Where motorcycle use is a small portion of traffic volume, such as the United States, some research has attempted to use motorcycle licensing and/or registration data as a surrogate, but these data are only available at the county level and so are not very useful for modeling site-level data.
Issues that apply to other rare crash types also hinder analyses of motorcycle crashes in order to develop SPFs and/or CMFs. Firstly, with rare crash types, low crash sample sizes make the development of reliable SPFs and CMFs difficult. Statistical models may not be estimable, or even if they are, estimated parameters may be very imprecise, and few variables related to crashes may be included in the model.
Another sample size issue is that roadway infrastructure treatments aimed at reducing motorcycle crashes are not common. When conducting a before-after study, the rarity of such treatment sites combined with low crash frequencies presents a formidable challenge.
The results of the research conducted for this project confirmed all of the analysis limitations stated above. Even with large datasets containing a substantial mileage of roads, SPFs could not be estimated for all site types or subtypes of motorcycle crashes. (Single-vehicle and multi-vehicle were attempted as well as total motorcycle crashes.) The simulation results estimating CMFs through before-after studies and cross-sectional regression modeling showed that the CMF estimates vary substantially between simulations due to low motorcycle crash frequencies.
Research Gaps and Needs
In terms of research gaps with respect to motorcycle safety and CMFs, very little information is known on the effects of roadway geometric and traffic control features on motorcycle crash frequency and severity. The reason for this gap is likely twofold: motorcycle crashes are not usually the focus of safety related countermeasures, and, the rarity of motorcycle crashes combined with the scarcity of treatment locations would result in a small sample size for study. Future research will need to explore how to overcome the small sample size issue with appropriate methodologies. This discussion provides some thoughts in this regard.
With respect to SPFs for applying in network screening and other safety management tasks, few SPFs at the segment level or intersection level exist. The SPFs developed in this project may contribute to filling this void, but there remains work to be done in terms of site types for which no SPF was developed and ensuring that models exist that calibrate well in all jurisdictions.
A major need for the research community would be a database that includes countermeasures implemented that are expected to affect motorcycle crashes along with the location, date of treatment, and treatment description. This information would aid researchers in identifying treatments that are feasible for study.
In terms of analytical methods and other related gaps, the project team identified of the following research needs.