validation of highway engineering data quality on ... 31, 2010 · validation of highway engineering...
TRANSCRIPT
Outline
• Introduction
• Project Goal
• Good Data Quality Characteristics
• Common Issues Leading to Poor Data Quality
• Research Methodology
• Results
• Conclusions/Recommendations
Introduction
• Anecdotal evidence & literature point to crash report interpretive errors being common place
• Quality data necessary for analysts to draw any meaningful conclusions
• Police officers have a limited background in highway engineering
– Training
– Instruction manuals
– Common sense
ehow.com
Project Goal
• Compare officer responses on MV4000 crash reports against researcher responses at the same location
• Determine where differences in data interpretation may require additional training or clarification
• Limited to highway engineering data fields
Good Data Quality Characteristics
• FHWA’s Crash Data Improvement Program Guide (CDIPG) published in April 2010
• According to CDIPG, crash databases should have the following “Six Pack” – Timeliness – Accuracy – Completeness – Consistency/Uniformity – Integration with other databases – Accessibility of data
• This project focused on accuracy, completeness, and consistency of engineering data fields of the Wisconsin crash database
safety.fhwa.dot.gov
Good Data Quality Characteristics
• Accuracy
– Internal validation (electronic forms)
– External validation (this research)
• Completeness—no “blank” entries
• Consistency—uniformity among files
– All WI agencies use MV4000
– National standard is Model Minimum Uniform Crash Criteria (MMUCC)
MMUCC
• Joint NHTSA & Governors’ Highway Safety Association publication
• Voluntary guideline, but compliance
necessary for States to receive Traffic Safety Information System improvement grants (Section 408)
• 2010 NHTSA Traffic Records
Assessment found WI MV4000 compliant in 90.6 percent of data fields and 53.1 percent in attributes – Field: Traffic Control – Attribute: Signal, Stop, Yield, etc.
People & Training
• According to the CDIPG, one refrain commonly heard from police is that “crash forms are being completed just for insurance companies”
• WisDOT MV4000 Instruction Manual
– Primary training resource for WI officers
– Last updated in 1998 – Brief and vague concerning
engineering fields • No baseline definition of when to flag
hills or curves • Poor definition of traffic barrier • No discussion of roundabouts
Processes
• Paper crash forms still used by approximately 45 percent of WI agencies in 2010
• Paper forms require manual data entry – Time consuming – Constrain resources – Lead to backlogs affecting
timeliness of data
• Personnel under pressure to process backlogs are prone to errors
Paper crash report backlogs in Texas. (GAO-10-454)
Technology
• CDPIG strongly recommends States adopt new and innovative technologies to improve crash data quality
• Two most commonly advocated technologies
– Electronic crash forms
– GPS based smart maps for location ID
Electronic Crash Forms
• Allow validity checks to catch inadvertent errors and blank fields
• Clean, clear crash diagramming
• Increases speed of submission, and therefore availability of data for analysis
• WI State Record planners predict 100 percent electronic reporting statewide by 2015
reportbeam.com
GPS/GIS Location Identification • MMUCC recommends that highway engineering data studied in this
research be generated by linking to roadway inventory data • Several States, including Maine and Ohio, already have such systems in
place • Badger TraCS (most common electronic report system in WI) has
interfaces for point-and-click location ID, but not yet used
Maine DOT Map Viewer System (GAO-10-454)
Other Crash Data Quality Studies
• Several studies have investigated commercial vehicle crash data quality, concluding that officer training needed improvement
• Mickee (2008) audited ~1000 Massachusetts crash reports – Used police, highway department & university auditors
– Reports were not verified by site visits
– Concluded that auditing created detailed statistical results, but was prohibitively time consuming
• Importantly, no other known studies incorporated investigation of verifiable conditions in the field
Research Methodology
• MV4000 data fields – Highway classification (derived based on municipality
population) – Access control – Traffic-way – Horizontal curvature – Vertical curvature – Traffic control
• Easily observable permanent highway features • Independent of other crash variables (e.g. date,
time, weather, driver, etc.)
Research Methodology
• Analysis timeframe: 1/1/09 to 12/31/10
• Geographic boundaries – Dane & Rock Counties – Representative of
municipality types in WI
• 27 intersection locations – Minimum of 5 crashes
2009-2010 – Apparent discrepancies
in three or more fields
• 664 total crash reports
Madison
Janesville
Beloit Bing.com
Site visit locations
Site Visits
• Observed ~250 foot radius around intersection
• Marked appropriate field attribute for each intersection approach
• In accordance with WisDOT Training Manual when possible—used engineering judgment when manual did not offer clear guidance – Curve: Alignment change of 15
degrees or more – Hill: Noticeable change in
elevation – Thinking from perspective of an
officer—they would not be measuring curve radii or grades
USH 51 at W Delavan Drive (Janesville)
Data Coding
• Binary system – 1 correct officer response
– 0 incorrect officer response
• Each crash report analyzed, using narratives and diagrams to determine correct approach where crash occurred
Field Report Coding
DOCTNMBR 1234567 N/A
HWYCLASS U CITY 1
ACSCNTL PART 0
ROADHOR C 1
ROADVERT 1
TRFCWAY ND 1
TRFCNTL1 TS OP 1
TRFCNTL2 NONE 0
Example of data coding for a report
Overall Results
• Highway classification – In general corresponded well – Exceptions often seen on
municipal boundaries (e.g. Fish Hatch at Greenway Cross)
• Traffic control – In general fairly accurate – Most common error was
reporting no control when in fact vehicles were subject to signals because of proximity to the intersection
• Access control, horizontal & vertical curvature, and traffic-way were more problematic
Correctly
Reported Percent of Total
TOTAL 664 100.0%
HWYCLASS 649 97.6%
ACSCNTL 480 72.3%
ROADHOR 602 90.8%
ROADVERT 613 92.9%
TRFCWAY 479 72.3%
TRFCNTL1 573 88.7%
TRFCNTL2 585 90.1%
Global Reporting Accuracy
Results: Access Control
• Overall accuracy only 72.3 percent • Researchers noted a wider variability in officer responses at locations
with at least one partially controlled facility (no private driveways, but at-grade intersections—e.g. USH 51 through Madison)
• Results confirm a sharp decrease in accuracy at partially controlled locations
Locations with Partial Access Control
Location
Reporting Accuracy (number
correct) Percentage
USH-14 & Pontiac 8 of 34 23.5%
STH-113 & CTH M 0 of 8 0%
US-51 & CTH CV 2 of 16 12.5%
W Main St. & O'Keeffe 2 of 13 15.4%
USH-51 & Pflaum 45 of 90 50.0%
USH-18 & CTH PD 22 of 50 44.0%
Total 79 of 211 37.4%
Locations with No Access Control
21 Remaining Locations
Reporting Accuracy (amount
correct) Percentage
Total 401 of 453 88.5%
USH 51 at Pflaum Rd (Madison)
Results: Horizontal Curvature
• Eleven sites had a curve present on at least one approach that should have been flagged
• Accuracy at locations with a curve present was much lower than those without a curve
Location Number of
Observations
ROADHOR
Accuracy
STH 26 & CTH N 23 69.6%
USH-14 & N Pontiac 34 97.0%
USH-51 & W Delavan 13 100.0%
Maple & Fourth 7 71.4%
US-51 & Anderson (CTH CV) 16 100.0%
W Main & O'Keeffe 12 91.7%
Century & Allen 14 78.6%
University & Allen 13 100.0%
Cumulative Total 132 89.3%
Other 16 Locations 481 97.1%
Results: Vertical Curvature
• Seven sites had a hill present on at least one approach that should have been flagged
• Accuracy at locations with a hill present was much lower than those without a hill
Location Number of
Observations
ROADVERT
Accuracy
STH 26 & CTH N 23 90.5%
USH-51 & W Delavan 13 84.6%
Johnson & Park 27 92.6%
Fish Hatchery & Post 18 72.2%
Fish Hatchery & Caddis 14 66.7%
US-14 & Deming Way 17 88.2%
University & Allen 13 38.5%
Cumulative Total 125 76.8%
Other 20 Locations 499 98.6%
Results: Traffic-way Divided without Barrier Totals Percentage
Correctly Reported 374 70.3%
Divided with Barrier 92 17.3%
Not Divided 59 11.1%
Blank 5 0.9%
One Way 2 0.4%
Total 532 100.0%
Not Divided Totals Percentage
Correctly Reported 18 41.9%
Divided without Barrier 24 55.8%
Blank 1 2.3%
Divided with Barrier 0 0.0%
One Way 0 0.0%
Total 43 100.0%
One Way Totals Percentage
Correctly Reported 36 97.3%
Not Divided 1 2.7%
Divided without Barrier 0 0.0%
Divided with Barrier 0 0.0%
Total 37 100.0%
True divided with barrier traffic-way http://forums.trinituner.com/upload/data/45/jersey%20barriers.jpg
• When facility was divided without a barrier, most common error was reporting a barrier • Roadways with no division (i.e. two lane road with painted centerline) were reported as divided over half the time
Results: Roundabouts
• Three roundabout intersections were studied
• The most inaccurate fields were traffic-way, horizontal curvature, and traffic control
S Towne &
Industrial
8th St &
Springdale
Commercial Ave
& N Thompson
Round-
about
Total
HWYCLASS 100.0% 100.0% 100.0% 100.0%
ACSCNTL 69.2% 100.0% 83.9% 82.0%
ROADHOR 30.8% 66.7% 29.0% 34.0%
ROADVERT 100.0% 100.0% 96.8% 98.0%
TRFCWAY 53.8% 50.0% 54.8% 54.0%
TRFCNTL1 92.3% 60.0% 86.2% 85.1%
TRFCNTL2 61.5% 100.0% 72.4% 72.3%
Results: Roundabouts
• Horizontal curvature reporting errors were tied to the circulating lane
• Majority of crashes flagged with a curve involved at least one vehicle in the circulating lane
Results: Roundabouts
• Traffic-way errors were most commonly associated with incorrectly using “not physically divided” or divided with a barrier
• Low traffic control accuracy most often resulted from officers indicating no traffic control, when the correct response should have been a yield sign
Results: Completeness & Agency Trends
• On the 664 crash reports in the study, there were only 16 “blank” entries (out of a possible 4,648)
• No notable trends were found based agency type (e.g. police, sheriff, state patrol), however, only 6 locations had multiple responding agencies
Conclusions
• Low accuracy for access control at partially controlled facilities indicates a lack of understanding by officers for what qualifies as partial control
• When hills or curves are present on at least one approach, officers are flagging them even when they are not the site of the crash, which is incorrect
• Low traffic-way accuracy is a result of misunderstanding what constitute divided roadways and barriers
• Roundabout-specific inaccuracies were especially noteworthy in the horizontal curve and traffic-way fields
Recommendations
• Update WisDOT Instruction Manual and develop pocket-sized “cheat sheet” versions
• In-service officer training with case-studies of commonly misinterpreted data fields
• Generate data fields by linking to roadway inventory and hardware data
Recommendations
• Implement 100 percent electronic crash reporting in Wisconsin
• Utilize GIS/GPS location identification technologies
• Update data fields to comply 100 percent with MMUCC, including roundabout-specific attributes