TEMPORARY LOSSES OF HIGHWAY CAPACITY AND IMPACTS ON PERFORMANCE: PHASE 2

Title | Contents | Acknowledgements | Exec. Summary
1. Intro | 2. Approach | 3. Crashes | 4. Breakdowns | 5. Work Zones | 6. Weather | 7. Signal Timing
8. RR Crossings | 9. Toll Facilities | 10. PUD | 11. Results Summary | 12. Next Steps | 13. References

3. Crashes

3.1 Methodology

The concepts described in Section 2 were used to estimate capacity reductions and resulting delays due to vehicle crashes.  As discussed before, crashes were defined by the following variables:  crash type (fatal, injury, property-damage-only [PDO]), number and type of vehicles involved, the location of the crash, the time of the day and day of the week the crash occurs, and the crash duration.  A three-step process was used for estimating delay from vehicle crashes:

Step 1. Vehicle crashes were assigned to the highway system.

Step 2. Capacity reductions were estimated based on crash type (fatal, injury, or PDO crash), number and type of vehicles involved, the location of the crash, the time of the day the crash occurs, and the crash duration.

Step 3. Delay was estimated based on capacity reduction, vehicle demand, the time of the day and day of the week the crash occurs, and the duration of the capacity reduction.

Due to differences in the characteristics of freeways and principal arterials, capacity reductions and delays for these two highway groups were calculated using slightly different data and methods.

3.1.1 Assigning Crashes to the Highway Network

The first step in estimating capacity loss was to assign crashes to a time and location within the national highway network.  Two primary data sources were used:  the General Estimates System (GES) compiled by the National Highway Traffic Safety Administration (NHTSA) and the Highway Performance Monitoring System (HPMS) maintained by the Federal Highway Administration (FHWA).

General Estimates System (GES):  GES data are collected from a nationally representative sample of police-reported crashes, both fatal and non-fatal.  To be eligible for the GES sample selection, a police accident report (PAR) must be completed for the crash; the crash must have involved at least one motor vehicle traveling on a highway; and the incident must have resulted in property damage, injury, or death. Data collectors for the GES make weekly visits to approximately 400 police jurisdictions in 60 sites across the United States.  They randomly sample about 50,000 PARs every year from these police jurisdictions. These collectors obtain copies of the PARs and send them to a central contractor for coding.  No other data is collected beyond the selected PARs—no driver license, vehicle registration, or medical information is obtained. The system began its operation in 1988.  At the time of the initial TLC study, the latest available GES data was for 1999.

Highway Performance Monitoring Systems (HPMS):  The HPMS, maintained by the FHWA, provides data that reflects the extent, condition, performance, use, and operating characteristics of the nation's highways (U.S. DOT/FHWA 2000).  Data collected for 1999 was used for this study.  HPMS data compilation is a cooperative effort between FHWA and state highway agencies, local governments, and metropolitan planning organizations (MPOs) working in partnership to collect, assemble, and report necessary information.  The physical and operational characteristics of highway facilities on which temporary capacity loss events occurred were drawn from information contained in the HPMS.

The GES data set was used to assign both fatal and non-fatal crashes to the highway system.  GES does not specify the exact location of the crash:  it merely specifies the region (Northeast, Mid-West, South, and West), land use type (large central city, suburb, other), general facility type (Interstate or non-Interstate), and number of lanes.  Therefore, a Monte Carlo simulation was used to place each crash on a surrogate highway location that would be similar to the environment under which the crash occurred, as described in the GES.

In the initial TLC, the Fatality Analysis Reporting System (FARS) compiled by NHTSA was used to assign fatal crashes to freeways, while GES was used to assign fatal crashes to principal arterials and to assign non-fatal crashes to both freeways and principal arterials. For TLC2, GES data was used to assign all crash types, including fatal crashes on freeways. Although FARS data for fatal crashes is typically believed to be more accurate, it was still necessary to make several assumptions regarding crash location, lane closures, and capacity reduction. Therefore, GES data was used so that a consistent data set, methodology, and assumptions could be used for all crash types. Also, any differences in data quality between GES and FARS should be relatively small compared to the potential error introduced by assumptions used in the methodology.

For crashes on Interstates, the region data was used to narrow down the possible states in which the crash may have taken place.  Then, the land use type and number of lanes from GES and the annual average daily traffic (AADT) from HPMS were used in the Monte Carlo simulation to place the crash on the freeway system.[2]  The process for facilities classified as Other Freeways and Expressways was similar. However, since crashes on these roadways are included in the GES "non-Interstate" category with all other non-Interstate highway crashes, a method was used to disaggregate these crashes to more-specific facility types.  In this study, crashes were "shared" to different non-Interstate highway types based on vehicle-miles of travel (VMT) by highway class as estimated in the VM-2 table in Highway Statistics (U.S. DOT/FHWA 2001). These crashes were then assigned to the highway network using the same method as for Interstates.

3.1.2 Estimating Capacity Reductions from Crashes

Capacity Reduction on Freeways

Crashes typically produce a loss of capacity on a freeway.  This capacity loss depends on the total number of lanes available and the number of lanes affected by the crash.  Table 7, adapted from Table 3, was used to determine such capacity losses.  For example, a crash causing the closure of one lane on a two-lane highway would reduce the total capacity available (i.e., 2 lanes × 2000 pcphpl) to approximately one-third of its normal value (i.e., 0.320 × 2 lanes × 2000 pcphpl) as opposed to one-half (i.e., 0.5 × 2 lanes × 2000 pcphpl) as a naive model would predict.[3]

The total number of lanes at the location where the crash occurred was obtained from GES.  However, data on the number of blocked lanes was not generally available.  Lane blockage depends on the type of crash and the number and type of vehicles involved.  It was assumed that a fatal or injury crash involving more than one vehicle always results in lane closures (i.e., probability of lane closures = 100 percent for injury crashes).

To estimate the probability of lane closures when only one vehicle was involved, the study proceeded as follows.  First, the probability of a fatal crash not causing lane closures was estimated.  From the FARS database for 1998, the total number of fatal crashes that were either located outside of the facility right of way or classified as off-road crashes (1,786 crashes) was estimated.[4]  This number was then divided by the total number of fatal crashes in the database for which the location of the crash was known (16,542 crashes out of 16,605).  This resulted in a probability of 0.108 (i.e., 1,786 ÷ 16,546 = 0.108) that no lanes would be closed due to the crash.  The probability of lane closures was then computed as 1 - 0.108 = 0.892.  For injury crashes (i.e., crashes resulting in non-fatal injuries), the same probability of lane closure as for fatal crashes was adopted.

Crash frequency information derived from Giuliano (1989) was used to determine the remaining probabilities of lane closures.  Utilizing these frequencies, non-injury crashes had a probability of 0.6 that lanes would be closed (and a complementary probability of 0.4 that no lanes would be closed), while breakdowns were less likely to close lanes (i.e., the probability was 0.154 that a lane would be closed due to a disablement).  For property-damage-only (PDO) crashes, these probabilities were overridden if more than three cars or more than one truck was involved in a crash.  This study assumed those crashes resulted in lane closures.  All fatal crashes involving multiple vehicles were assumed to close lanes.  The probability of a fatal crash involving a single vehicle closing a lane was assumed the same as for injury crashes closing a lane.  Table 8 summarizes lane closure probabilities.

Table 9 shows the number of lanes closed due to the crash as a function of the number and type of vehicles involved, which is used as a proxy to describe the severity of the crash.  It is assumed that, at most, four lanes could be closed due to the crash.  This assumption was made since extending the information on capacity losses to cases with more than four lanes closed (see Table 7) could be highly unreliable without further studies (using simulation, for example).

For crashes involving a large number of vehicles (more than 3) and occurring on facilities with more than 4 lanes, this assumption may underestimate the delay caused by the crash since more than four lanes could actually be closed.  Moreover, the number of lanes closed presented in Table 9 should be a function of the type of crash, since even those crashes involving only one vehicle may result in the closure of all lanes (e.g., a hazardous material spill).  In an attempt to capture these types of occurrences, a probability slightly larger than zero (0.001) was given to the closure of any number of lanes for any number and type of vehicles involved.

Capacity Reduction on Principal Arterials

The procedure used for predicting capacity losses on arterials due to crashes was similar to the one described above for freeways, although some elements differ.  Due to a lack of better information, it was assumed that the capacity losses on principal arterials were the same as for crashes on freeways. Thus, it was assumed that the capacity reduction values in Table 5 were also valid for principal arterials.

As explained before, the next step was to determine whether there were lane closures due to the crash. The location distribution for freeways was used as a surrogate for principal arterials. However, since most arterials do not have a shoulder, it was assumed that any crash would produce a lane closure, independent of the type of crash and the number and type of vehicles involved.

Due to a lack of information, it was assumed that the number of lanes closed was the same as for freeways (see Table 9).  It was also assumed, however, that a severe crash on a principal arterial would likely close lanes in both directions of traffic.  To account for this, the total number of lanes in both directions was considered when assigning the number of lanes closed.  For example, if the model indicated that a crash under consideration closed four lanes on a principal arterial comprised of three lanes in each direction, it was assumed that one direction would be completely closed to traffic, and one lane in the opposite direction would be closed.

3.1.3 Estimating Delays Due to Crashes

Delays on Freeways

Delays were calculated as described in Section 2, "Approach and General Methodology."  Three important variables were involved in this computation:  (1) the time of the day and day of the week when the crash occurred, (2) its location, and (3) its duration.

Location (rural or urban) together with time of day and day of week were used to determine the demand on the freeway during the interval of time the roadway was affected by the crash.  This was done by adjusting the AADT for that location through multipliers obtained from the appropriate demand distribution curve. The study used demand distributions for day of the week and time of the day derived from information collected for four cities:  San Antonio, Texas; Milwaukee, Wisconsin; San Diego, California; and Seattle, Washington. [5]  For rural areas, information collected for the state of Tennessee was used. These distributions are shown in the figure below.

Fig. 10. Traffic volume distributions for urban and rural areas.

The first step in determining the demand was to adjust the AADT for heavy vehicles.  Due to lack of better information, it was assumed that each crash happened on a segment of freeway 0.5 to 1 mile long with a grade of 1 percent.  This assumption resulted in a passenger car equivalent (PCE) of 3 for any one percent of trucks in the traffic stream.  Information about annual average daily truck volumes (AADTT) on the facility where the crash occurred was obtained from HPMS.  To obtain the PCE demand (call it AADT'), the AADT was adjusted by multiplying the estimated number of trucks by three and treating them as passenger vehicles. AADT already contains AADTT, and since the PCE for heavy vehicles was determined to be 3, then AADT' = AADT + 2 × AADTT.  The AADT' was then multiplied by 7 to obtain the total volume for the week.  The result was then multiplied by the factors (obtained from the urban or rural demand distributions) corresponding to the day of the week and hour of the day when the crash occurred.

Crashes were assumed to be composed of three intervals:  (1) the crash detection/arrival-to-scene interval, (2) the remove-to-shoulder interval, and (3) the clearance interval.  The crash detection/arrival-to-scene interval is the time that elapses between the actual occurrence of the crash and the time at which the corresponding emergency management (EM) personnel arrive at the location of the crash.  During this interval, it was assumed that the vehicles involved were blocking a number of lanes determined using the procedure explained above.  For a large percentage of the fatal crashes, the database contained information on the time at which the crash occurred and the time at which the police or other EM personnel arrived at the scene.  For those cases, the crash detection/arrival-to-scene time was computed as the difference (in minutes) between these two time points.  If any or both of these times were not known, a detection/arrival-to-scene time of 10 minutes was assumed, which is slightly larger than the one reported by Skabardonis et al. (1998).

The remove-to-shoulder interval (RSI) represents the time required to move the vehicles from the roadway to the shoulder.  Table 10 shows the remove-to-shoulder duration times as a function of the number and type of vehicles involved in the crash.  During this interval, it was assumed that the vehicles involved were blocking a number of lanes determined using the procedure explained above.

The crash duration (CD) time represents the time elapsed between the arrival of emergency management personnel to the scene and the time at which the crash was totally cleared (this comprises both the remove-to-shoulder time and the clearance time).  The TLC study used information derived from Giuliano (Table 11).

Knowing the type of crash and whether or not the crash resulted in lane closure (which was previously determined to assess the loss of capacity), the table gives the mean of the crash duration and its standard deviation.  Table 12 below, derived from Table 11, shows these mean duration times.  (Note:  only the means are used in this version of this study.)

The remove-to-shoulder interval (RSI) is assumed part of the CD interval.  That is, for those cases where RSI <=15 minutes, it was assumed that, during the interval CD - RSI, the vehicles involved in the crash where located on the shoulder of the freeway, producing a capacity loss obtained from the first row of Table 7.  Otherwise, it was assumed that, during the entire interval CD, the vehicles were blocking 1 or 2 lanes, as determined above, producing the corresponding capacity loss from Table 7.

Delays on Principal Arterials

Delay on principal arterials was estimated using the same method used for freeways, although additional assumptions were necessary.  Principal arterial traffic demand was adjusted using the multipliers for time of day and day of the week, just as it was for freeways.  Due to the lack of better information, the same multipliers used for freeway demand were used for principal arterials.

Since it was assumed that principal arterials do not have shoulders, crashes on those types of facilities were composed of just one interval, rather than the three intervals used for crashes on freeways.  In most cases, the length of this interval was assumed equal to the duration of the crash (see Table 12) plus another 10 minutes to account for the detection of the crash.  However, there is one exception to this rule.  As previously stated, the freeway lane-location distribution was used to locate non-fatal crashes on principal arterials.  However, "property damage only" (PDO) crashes (i.e., non-fatal, non-injury) involving less than four vehicles and assumed to end up on the shoulder in the location distribution were assigned a much smaller duration.  A duration of 15 minutes was assumed for this kind of crash (10 minutes for the arrival of police to the scene and another 5 minutes to move the vehicles out of the way).

The delays were calculated using essentially the same approach used for freeways, with the following differences:  First, an ideal capacity of 1,600 vphpl was assumed.  Second, the green time percentage for the principal arterial was generated from a uniform distribution with lower and upper bounds of 50 and 70 percent, respectively. [6]  Since a principal arterial should get a green light at least half of the time, a lower bound of 50 percent was assumed.  The upper bound of 70 percent green time was also based on assumption.  Third, it was assumed that the principal arterials form a grid with separation L uniformly distributed in the interval [0.5, 3.0] miles.  It was further assumed that the crash had the same probability of being located anywhere along the arterial under consideration.  Calling d the distance from the immediate upstream transversal arterial to the location at which the crash occurred (d <= L), it was assumed that the queue due to the crash could not be longer than d, since traffic would likely divert at the transversal arterial upstream of the crash.  In effect, the size of each queue on an arterial was limited by truncating it at a length equal to d.  This represents traffic diversion in an arterial network grid.  Although the traffic diverting would not experience any delay due to the crash, the vehicles would incur a longer trip and, in consequence, longer travel times.  However, since the effects of re-routing were beyond the scope of the present study, those longer travel times were not computed as part of the delays due to crashes on principal arterials.

3.2  Results

The TLC2 study estimates that, in 1999, approximately 3.3 million crashes on freeways and principal arterials caused temporary capacity reductions of about 3.3 billion vehicles and over 1.7 billion vehicle-hours of delay (Fig. 11, Table 13).  Crashes on freeways caused nearly three times as much delay as those on principal arterials. Freeway crashes produced around 1.2 billion vehicle-hours of delay (73 percent), while crashes on principal arterials accounted for around 459 million (27 percent). The average delay caused by each freeway crash (1.2 thousand vehicle-hours) was about six times more than the average for each principal arterial crash (197.5 vehicle-hours).  Similarly, average delay per crash was 50 percent higher for fatal crashes than for non-fatal crashes. Still, due to the higher frequency of non-fatal crashes and the fact that fatal crashes typically occur during non-congested conditions, non-fatal crashes accounted for about 99 percent of delay from all crashes.

Figure 11.  Delay per crash by severity and highway type

3.3 Reliability

3.3.1 Methodology

The methodology can be divided into three parts: assigning crashes to the network using a Monte Carlo simulation, estimating capacity loss, and estimating delay. 

It is believed that the Monte Carlo simulation method described in section 3.1.1 was a fairly reliable method for assigning crashes to highway links similar to those on which the crash actually occurred. As a limited test of variability, capacity reduction and delay were generated for two datasets, each consisting of ten thousand assigned crash locations. The results for these two data sets were similar. While this is a limited indicator of reliability, it is encouraging.

To estimate capacity loss, capacity reduction multipliers were obtained from the literature or estimated when not available―the reliability of these data and assumptions are discussed below. To estimate delay, the TLC study used standard, well-established procedures and methods derived from traffic flow theory as presented in the Highway Capacity Manual (TRB 1985).

The confidence level of the methodology used here to determine capacity losses and delays due to fatal crashes depends on the type of facility where the accident occurred.  For freeways with up to five lanes where at most two lanes are closed, both methodologies can be classified as having a high degree of confidence.  For freeways with more lanes (either total and/or closed), the computation of the delays retains a high level of confidence, but the capacity reduction estimates become more unreliable since they are outside the boundaries of the studies from which this information was collected. For arterials and other facilities with unrestricted access, both methodologies can be qualified as having a medium degree of confidence.

3.3.2 Data and Key Assumptions

Crash Locations and Characteristics: The General Estimates System (GES) data used to assign crashes and crash characteristics to the highway network is based on a random sample of about 50,000 PARs every year from about 400 police jurisdictions. It is widely used as a data source for crash-related studies and is qualified as having a high degree of confidence.

Traffic Demand: The traffic demand data is accorded a low level of confidence since it was derived using Annual Average Daily Traffic (AADT) information from HPMS along with time of day and day of week distributions based on freeway data from only four cities and one state.  Because of the high impact that demand can have on the computation of delays, further research in this area would need to be performed in future versions of this study.

Probability of Lane Closure:  The data and assumptions used for determining lane closure probabilities are classified as having a high degree of confidence for freeway crashes but a lower degree of confidence for arterial crashes.

Number of Lanes Blocked by the Event:  The probability distribution of the number of lanes closed by severity of the crash was created based on engineering judgment—number and type of vehicles involved in the crash were used as a proxy to describe its severity. Due to the lack of resources to corroborate the resulting distribution, the data is assigned a medium to low degree of confidence.

Surface Street Characteristics: For principal arterials, the ideal lane capacity, the traffic signal green ratio along the arterial, the geometry of the network (i.e., the size of the arterial grid or separation between transversal major arterials), and the location of the crash along the arterial segment in relationship to the closest upstream major transversal arterial were based on engineering judgment.  These assumptions can be qualified as having a medium degree of confidence.

Duration of the Event: Crash durations were based on detection and clearance times obtained from several papers that studied freeway incidents, as well as assumptions based on engineering judgment. Assumptions regarding crash duration are accorded a high degree of confidence for freeway crashes. They are accorded a medium to low level for arterial crashes since the studies on which the assumptions were based included freeways only.

2. The average of the AADT for the freeways was taken for the freeways matching the land use and number of lanes. The Monte Carlo simulation places the incident on a freeway segment that has an AADT close to the average.

3. pcphpl = passenger car equivalent (PCE) vehicles per hour per lane. Passenger car equivalent is a measure used to account for the different size and performance of most multi-axle vehicles.

4. Although FARS was not used to assign crash locations in TLC2, the data was used to derive the probability of a vehicle involved in a crash to leave the roadway and not block lanes.

5. Data compiled by Oak Ridge National Laboratory (ORNL) from real-time traffic count data from San Antonio, Texas; Milwaukee, Wisconsin; San Diego, California; and Seattle, Washington. For more information see Chin et al., 2000; Chin et al., 1999.

6. Since the ideal capacity represents the flow on uncontrolled links, the green time percentage is used to adjust the capacity to account for the fact that traffic is stopped part of the time.

Previous | Table of Contents | Next