Crash Data Collection and Quality. Why collect/maintain safety data? Khisty says: Is that all? –...

Crash Data Collection and Quality

Why collect/maintain safety data?

• Khisty says:

• Is that all? – Better understanding of operational problems– Accurate diagnosis of crash problems– Develop remedial measures– Evaluate the effectiveness of road safety programs

Who uses crash data?– Road safety engineers

• Develop remedial measures

– Police• Charging a person at fault in

crash• Enforcement activities

– Location of speed cameras– Breath testing stations

– Insurers• Seeking facts before settling

claims

– Lawyers• Compensation for injuries

– Road safety educators• To ensure that their efforts

well targeted

– Safety administrators• Report statistical information

on road crashes

– Researchers• Access good reliable database

– Vehicle manufacturers• Assess the safety of their

products

Importance of good data (Video “L”)

And, for Commercial Motor Carriers …• Identifying the appropriate Commercial Motor Carrier• Determining Reportable Crashes• Identifying Vehicle Configuration and Cargo Body Type• Determining Sequence of Events• Recording Hazardous Materials• Recording proper CDL

Supplementary data sources (Ogden)

• While police crash report is the basic source of crash data, there are some other sources which may be useful and applicable in certain circumstances– Local knowledge

• Local government staff• Emergency service personnel• Local safety groups• Local businesses

– Interview of road users• People involved in a crash at a site of interest, which are source of useful

information for traffic officials in development of countermeasures– In-depth studies of particular group of crashes

• Single vehicle fatal crashes, to gain better understanding of the nature of those crashes

Supplementary data sources (continued)

– Traffic conflict surveys • May be used when the collection of crash data is not practical or period

of evaluation is too short to collect sufficient samples– Field observation– Video recording of conflicts

• Information gained in this way is valuable in – getting a sound understanding of the traffic operation– Find interactions between traffic streams at the site

• As a proxy measure of safety– Assumption must be made about relationship between proxy measure

(conflict) and crash rates

– Site investigations are necessary component of a countermeasure development program

What is reportable?• In most US states, the five point scale often referred to as

KABCO– K person with fatal injury– A person with incapacitating injury– B person with non-incapacitating evident injury– C person with possible injury– O no injury (property damage only)

• Some countries report injury crashes only• Some states do not differentiate between injury types

– Implication?

• Some crashes are not reported … why?• Many states use a reporting “threshold”

– May vary even within states … implication?

Impact of threshold adjustments

Sketch and narrative

Collision Diagrams

http://www.nhtsa-tsis.net/stateCatalog/stateData.html

In-class exercise

Crash form elements and the Haddon Matrix

Storage/retrieval

• <500 annually may be filed (paper) with summary tables

• Increasingly, all data are input into a database (and forms scanned)

• Feeds state and national databases

Old Location Process

Data Collection Technologies

• TraCS: Traffic and Criminal Software

TraCS data entry form

Incident Location Tool (and IMAT)

Easy Street Draw & Visio

Florida TraCS show

Case study – access management

From …

Use and Abuse of Crash Data in Roadway Access Management

A Workshop at the National Access Management Conference

Baltimore, Maryland July 13, 2008

Case study – access managementFrom …

Data-Driven Access Management• Access management treatments and plans should be directly tied

to measurable objectives such as crash or crash cost reduction• Access management treatments proposed should be appropriate

given the types of crashes and pattern of crashes being experienced in a corridor

• Access management treatment costs need to be justifiable based upon the expected benefits of crash reductions and other objectives • Stakeholders and decision-makers must be convinced that the “gain”

of access management is worth the “pain” • Confidence in both past (“before treatment”) and expected future

crash rates (“after treatment”) should be high• You want to be very sure that any treatments

will produce a noticeable and positive result

30

Access Management and Safety• Most access-management related

crashes occur on urban and suburban arterial roadways at speeds of 35 to 55 miles per hour

• Up to half of all crashes in urban areas are related to issues of access (minor public road intersections, traffic signal spacing, driveways)

• Although most access-related crashes occur in urban or suburban areas, access-related crashes in rural areas tend to be severe crashes due to higher travel speeds

• Access-related crashes occur at conflict points

• The diagram represents one crash data point

31

Problem 1: Fix This Mess South Ankeny Blvd., Ankeny, Iowa

32

What Do Crash Data Really Look Like?

33

What’s On Your Table …

34

Land Use

Crash data tables and charts

Crash data stack mapLaminated base map

Traffic over time

Corridor photos

34

An Example Plan …

35

Crash Data Allow Better …

• Problem Identification• Understanding of the problem before

jumping into exploring and designing solutions

• Focus on severe crashes rather than all (minor) crashes

However …

36

You Need Good Quality Data

The Ingredients Matter: Quality Control37

The Characteristics of Data Quality (The “Six-Pack”)

38

FMCSA Data Quality

Crash Data Quality: Timeliness• Sometimes crash data are not available for months or

even years• Varying timeliness of different jurisdictions can cause

issues for comparative analysis• Time itself is important – did something change

during the analysis period?• Also – the time period is important … one year of data

are probably not enough!

40

Crash Data Quality: Accuracy

• Spatial Location• Attributes, e.g.,

severity, crash type, roadway info

41SO

UTH

AN

KEN

Y BO

ULE

VARD

SOU

TH A

NKE

NY

BOU

LEVA

RD

11STST Road Roadv

OriginalOriginal

Considering Considering functional functional areaarea

Crash Data Quality: Completeness• Missing data can lead to a misleading

picture and erroneous conclusions• Some crash records have “unknown” or

“other” fields• Some crash records are missing altogether• Variations between jurisdictions (county

level, state level) can lead to inaccuracies in comparative analysis

• Random bias - Under-reporting can result in distorted picture of road crash situation

42

Collision TypeNum of Crashes Percentage

Non-collision 17854 32.6%

Head-on 1006 1.8%

Rear-end 12143 22.2%Angle, oncoming left turn 3528 6.4%

Broadside 10192 18.6%Sideswipe, same direction 5035 9.2%Sideswipe, opposite direction 1145 2.1%

Unknown 3538 6.5%

Not Reported 374 0.7%

Total 54815 100.0%

Crash Data Quality: Consistency/Uniformity

• Across jurisdictions• Across time• Consistent severities• Discontinuities

– Data from one time period can not be compared to another time period

43

Crash Vehicle Person Roadway

MMUCC and MIREModel Minimum Uniform Crash CriteriaModel Inventory of Roadway Elements

Consistency

• Although the need for data is universally recognized, there is little consistency in collected data (Ogden)– Comparative study of eleven European countries found

that• Only two variables (date & hour) were collected in all eleven

countries• 7 percent of items were recorded in three countries• 70 percent recorded in only one country

– There is no nationwide crash data reporting system in US• Little consistency within states for recorded data elements

Crash Data Quality: Integration

• Integration provides a ‘richer’, more complete source of information (e.g., integration with roadway features)

• Double check on accuracy (including severity)• Privacy is a tough issue• Another tough issue is multiple offices and even

agencies being in charge of various parts of safety data

46

Crash Data Quality: Accessibility

• How can you get crash data?• How easy is it to get?• What form do you want it in?• Liability and perception is an issue.• Continuum:

not available … special request w/delay … regular updates … service … instant web access

47

Typical Crash Data IssuesThese may not be apparent to the data user

Changes in Crash Forms• Content

– Addition/elimination of attributes collected– Change in definitions (values)

Non-collision

Head-on

Rear-end

Angle, oncoming left turn

Broadside

Sideswipe, same direction

Sideswipe, opposite direction

Head-on

Broadside/Left Turn

Rear End

Rear End/Right Turn

Rear End/Left Turn

Sideswipe/Opposite Direction

Sideswipe/Same Direction

Sideswipe/Right Turn

Sideswipe/Left Turn

Sideswipe/Dual Left Turn

Sideswipe/Dual Right Turn

Broadside/Right Angle

Broadside/Right Entering

Broadside/Left Entering

Head-on/Left Entering

Sideswipe/Both Left Turning

Single

Pedestrian

Bicycle

Parked Vehicle

Before After

Collision Type

49

Changes in Crash Forms, cont.Impacts:• Difficult to perform direct comparisons over analysis

period.• May result in systematic change in apparent crash

performance, e.g. crash reduction.

Year

Cras

h Ra

te

StatewideYear

Cras

h Ra

teSite #1

Change in crash form

50

Cartographic (Base Map) Changes

• Shift, update to reference road network

Impact: Challenging to systematically assign crash location.

51

Location Accuracy

• How are the crashes located? – GPS (where?)– Manually derived, based on literal description– LRS, Link-node, other?

• What reference networks are used?– GIS– LRS– Link-node

52

Location Accuracy, cont.

• How do accuracies vary among location methods and reference networks?– Ex. GPS ±5m v. GIS-based road network ±10m

Impact: type I or type II errors – you’d not know

X

Actual crash location

Crash may be locatedanywhere within this area.

Roadway may be presentedanywhere within this area.

X

Geocoded crash location

GIS road network

53

Changes in Statute

• Reportable crash definition– Property damage threshold, e.g. $500 v. $1000– Injury crash

• Reporting requirements– Driver report “…is not required when the accident

is investigated by a law enforcement agency.”

Impact: May result in systematic change in apparent crash performance, e.g. crash reduction.

54

Reporting Extent & Completeness

• All public roads• Private property• State-maintained roads only• Jurisdiction, agency dependent

Impacts:• Incomplete crash history skews findings.• Difficult to compare different locations.

55

Multiple Data Sources

• Local law enforcement• State DOT• Other agencies, e.g. taxi authority

Impact: Difficult to access and integrate all crash data, i.e. difficult to create a comprehensive, useable data set.

56

How Crash Data Are Abused

57

Limited Frame of Reference

• Limited, no comparison to similar locations.• No comparison to “expected” conditions

(comparables).

Impact:• What may appear to be a problem site, in isolation,

may be performing as well as, or better than, similar locations.– However, this does not imply that a location is performing

well and/or can not be improved.

58

Limited Perspective• Decisions made, almost exclusively, based on crash history.• Little consideration given to

changes during analysisperiod…– Land use and development– Infrastructure– Traffic patterns– Other, e.g. construction

during an analysis year

Impact: • Factors significantly impacting

crash history are ignored.• Solution no longer fits the

problem59

Regression to the Mean

• Crashes are random.• Extreme conditions will generally return to

“normal” state.

Source: Safe Speed Source: Safe Speed

Impact: Overestimates effectiveness of treatment; focus on the wrong sites (should use EB or at least more data) 60

Analysis Period Shortcomings

• Limited (short) analysis period • “Dated” crash data

Impacts:• May not accurately represent the performance of a

site. Similar to regression to the mean.• May not accurately reflect the existing conditions.

61

General Crash Data Issues

• Change in crash form• Cartographic (base

map) changes• Location accuracy

• Change in statute• Reporting extent &

completeness• Multiple data

sources

Impact: Not being aware of these issues – is it your responsibility?

62

Problem 2: Fix This Mess Lincoln Way, Ames, Iowa

63

Data On Your Tables …

1. Complete set of data2. 25 meter buffer vs. “Functional area”3. Crash frequency only vs. AADT and crash

type4. 1 year of data vs. 10 years of data5. Older data vs. recent data6. Current aerial photo only vs. past

development trend and detailed land use data

64

Locational Challenges for Next Generation of Crash Data

Systems

SAFETEA-LU Section 1401 (Highway Safety Improvement Program)

ID of top 5% of public hazardous locations on all roads

Local Road GIS Data

Where some states are now

Inventory data on all roads?

The “quadrennial needs” legacy

YesSome, quality issue, or working on itNoNo Response

State system as a percent of all public roads

Can 1401 be met without GIS?

Kansas, for example …• Has crashes on system only • Has ≈ 70% of crashes located to

road by route milepost• Does sliding spot (nongraphical)

& “named intersection” (program)

• Assuming the 30% missing does not affect the outcome

• No brainer to do top 5%

Location• An early computerized “spot” map (from Khisty)

Can you “spot” the problems?

Other examples

Crashes by Time of Day

Crashes by Age

Crashes by Road Surface Conditions

Drug and Alcohol Related Crashes

GIS-ALAS: Corridor Crash Frequency (stacked)

Injury Frequency by Severity

Injury Frequency by Severity

Mason City

Waterloo

Cedar Rapids

Quad CitiesDes Moines

Council Bluffs

IowaCity

Ames

SiouxCity

DubuqueFort Dodge

Ottumwa

Marshalltown

Spencer

Clinton

1 yrof data

Crash Density – 1 Year AverageAnnual Fatal and Major Injury Crashes Per Mile

Sample

- DRAFT

Sample

- DRAFT

Mason City

Waterloo

Cedar Rapids

Quad CitiesDes Moines

Council Bluffs

IowaCity

Ames

SiouxCity

DubuqueFort Dodge

Ottumwa

Marshalltown

Spencer

Clinton

3 yrsof data

Crash Density – 3 Year AverageAnnual Fatal and Major Injury Crashes Per Mile

Sample

- DRAFT

Sample

- DRAFT

Mason City

Waterloo

Cedar Rapids

Quad CitiesDes Moines

Council Bluffs

IowaCity

Ames

SiouxCity

DubuqueFort Dodge

Ottumwa

Marshalltown

Spencer

Clinton

5 yrsof data

Crash Density – 5 Year AverageAnnual Fatal and Major Injury Crashes Per Mile

Sample

- DRAFT

Sample

- DRAFT

Mason City

Waterloo

Cedar Rapids

Quad CitiesDes Moines

Council Bluffs

IowaCity

Ames

SiouxCity

DubuqueFort Dodge

Ottumwa

Marshalltown

Spencer

Clinton

10 yrsof data

Crash Density – 10 Year AverageAnnual Fatal and Major Injury Crashes Per Mile

Sample

- DRAFT

Sample

- DRAFT

Location methods• address• offset from known point

(intersection, bridge, crossing, milepost)

• Literal description• Smart map• Lat/long or other coordinates (GPS)• Aerial photo

Multiple methods required

Spatial analysis methods

• Spot/Intersection Analysis • Strip Analysis • Cluster Analysis • Sliding-Scale Analysis • Corridor Analysis

Spatial statistics is an emerging area

But …some technical issues

Some not-so “simple” questions

Feature not represented

Feature under

construction

Alignment OK

Alignment Off

Where are the roads? (Incorrect or incomplete cartography)

Where are the roads? (Improving cartography)

Alignment moves

Alignment stays put

Where are the crashes?• Crashes are not

necessarily point events• Some crashes may be

located using different methods and degree of accuracy – Temporal (e.g. link node

to lat long)– Spatial (e.g., state police

v. local)– Techno (GPS v. smart

map)

?

What’s “the” traffic volume on “the” road?

• Need traffic level for the year the crash happened

• Requires multiple files – in Iowa, working on going back past 1998 – difficult to do

• Was the road even there then? Is the road still there?

How to segment the road system?

• Requirements– Logical breaks (engineering and

public)– Relationship to inventory data– Long enough for manageability

and presentation– Short enough to reflect

important changes– Clear and understandable to use

• Facility location and type– What is rural/urban? Character is

important …Designated

“rural”

Can use attributes and/or proximity…Red: probable, Yellow: spatial @ 75’, Blue: possible + spatial

What is an intersection crash?