People Saving People
National Center for Statistics & Analysis
Estimating Alcohol Involvement
in NHTSA’s
Fatality Analysis Reporting System (FARS)
Rajesh (Raj) Subramanian
Mathematical Analysis Division
29th Annual Traffic Records Forum, Denver, CO 1
Definitions
People Saving People
National Center for Statistics & Analysis
Blood Alcohol Concentration (BAC)
- Measure of concentration of Alcohol (grams) in blood
(deciliter).
Alcohol Related Fatality
- A fatality that occurs in a crash where at least one of the
involved drivers,pedestrians or pedalcyclist has a BAC of 0.01
or greater.
- BAC is the only metric used to determine this.
Intoxication
- In most states, a person with a BAC of 0.08 or above.
29th Annual Traffic Records Forum, Denver, CO 2
Under-reporting of BAC values
People Saving People
in FARS – A Historic Perspective
National Center for Statistics & Analysis
Proportion of Missing BAC Among
Actively Involved Persons – Drivers and Pedestrians
80
75
Percent Missing
70
65
60
55
50
45
40
35
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
Driver Pedestrian
29th Annual Traffic Records Forum, Denver, CO 3
Problems Posed by
People Saving People
Missing BAC
National Center for Statistics & Analysis
Hypothetical FARS Case
Vehicle 1 Vehicle 2
Police has indicated
Surviving Driver Fatally Injured
that there is
BAC=0 Driver (BAC=?)
probable cause to
Fatally Injured Fatally Injured
test driver of Vehicle 2
Passenger Passenger
for alcohol!
Are the three fatalities
in this crash
Alcohol-Related?
Cannot be Determined unless BAC is Estimated (Imputed)!
29th Annual Traffic Records Forum, Denver, CO 4
Imputation of Missing BAC
People Saving People
National Center for Statistics & Analysis
Estimate BAC only when it is missing
- For any driver, pedestrian or pedalcyclist with missing
BAC.
Estimation is done based on crash, driver and vehicle
related characteristics
- Various predictors of alcohol are used.
- Police-reported alcohol involvement (DRINKING) is
highly predictive of presence/absence of alcohol
Imputes actual value of BAC
- Instead of a binary response (Yes/No).
29th Annual Traffic Records Forum, Denver, CO 5
Candidates for Imputation
People Saving People
National Center for Statistics & Analysis
Involved Driver with Missing BAC value.
- Surviving or fatally injured Driver
Involved Non-occupant with Missing BAC value
- Surviving or fatally injured Pedestrian or
Pedalcyclist
29th Annual Traffic Records Forum, Denver, CO 6
FARS Variables used to Impute
People Saving People
Missing BAC
National Center for Statistics & Analysis
Variables Used
Police Reported Drinking License Status
Age category Previous Incidents (DWI, etc.)
Gender Day of the Week
Use of restraint Time of the Day
Injury severity Vehicle Role
Relation to Roadway
29th Annual Traffic Records Forum, Denver, CO 7
Overview of Imputation Process
People Saving People
National Center for Statistics & Analysis
Step 1
- Choose set of variables that are significant in predicting
dichotomous BAC (BAC=0 vs. BAC0)
Step 2
- Conditional on case having non-zero dichotomous BAC,
choose set of variables that are significant in predicting
continuous BAC.
Step 3
- Combine the results from Steps 1 and 2 into one general
model and impute missing BAC.
29th Annual Traffic Records Forum, Denver, CO 8
Imputation Domains and
People Saving People
Process Flow
National Center for Statistics & Analysis
Steps 1 to 3 are performed within each vehicle class
- Broad based vehicle categories like Cars, Utility Vehicles,
Other LTVs, Minivans, Medium and Heavy Trucks, Motorcycles
and all other vehicles.
- Characteristics of Drivers for one class of vehicles may be
radically different from those of another vehicle class (e.g.
Minivans vs. Motorcycles)
Non-occupants are treated as a separate class
- Lesser number of predictor variables (No license status,
restraint use, etc.).
29th Annual Traffic Records Forum, Denver, CO 9
Validation of Imputed BAC Values
People Saving People
National Center for Statistics & Analysis
Impute BAC values intentionally set to Missing
- Randomly set 25 percent of reported values to
missing and impute these “missing” values
- Compare imputed BAC values with the actual
reported values
- Repeat this test.
Analysis on Imputation and probable cause for testing
alcohol as indicated by Police Officer.
- Suspected alcohol involvement, or the lack thereof,
by the police officer.
29th Annual Traffic Records Forum, Denver, CO 10
Does it Work?… Validation Tests on
Estimating Alcohol-Related Fatalities
People Saving People
National Center for Statistics & Analysis
Rate of Driver and Non-occupant Randomly Set 25% of Known BAC
Alcohol Testing for USA, 2001 Values to missing and Impute
Known
= 44%
Randomly Set 25%
75%
Unknown to Missing
= 56%
29th Annual Traffic Records Forum, Denver, CO 11
25% Validation test on Alcohol-
Related Fatalities, 2000
People Saving People
National Center for Statistics & Analysis
Alcohol Related Fatalities as Percent of all fatalities
in crashes involving persons with known BAC values
Actual Values, 2000 25% test, 2000
Alcohol Related Alcohol Related
No Alcohol No Alcohol
Related 44% Related 44%
56% 11,257 56% 11,245
14,473
14,461
29th Annual Traffic Records Forum, Denver, CO 12
Analysis on Police-Reported Alcohol
People Saving People
Involvement and Missing BAC
National Center for Statistics & Analysis
Alcohol Test Results changed from unknown value in
the Annual Report File (ARF) to a reported value in the
FINAL file.
- Due to delay in test results being reported to FARS
- Missing BAC values were imputed in the ARF.
BAC values were reported for same cases in FINAL file.
- Presents an opportunity to assess imputed BAC
values.
29th Annual Traffic Records Forum, Denver, CO 13
Characteristics of Converted Cases –
People Saving People
Police Reported Alcohol Involvement
National Center for Statistics & Analysis
Of all the converted cases
- 35 percent indicated NO for police reported alcohol
involvement (DRINKING)
- 19 percent indicated YES for DRINKING.
- 18 percent were coded as NOT REPORTED.
- 29 percent were coded as UNKNOWN.
29th Annual Traffic Records Forum, Denver, CO 14
Comparison of Reported and Imputed
People Saving People
values for DRINKING=NO
National Center for Statistics & Analysis
- 95 percent of the FINAL reported BAC values were 0
- 91 percent of the imputed BAC values in the ARF
were 0.
- Reflecting an accuracy of close to 96 percent for this
group.
- Small proportion of cases ended up with contrary
result.
29th Annual Traffic Records Forum, Denver, CO 15
Comparison of Reported and Imputed
People Saving People
values for DRINKING=YES
National Center for Statistics & Analysis
- 88 percent of the FINAL reported BAC values were
greater than zero.
- 84 percent of the imputed BAC values in the ARF
were greater than zero.
- Highlights importance of Police-Reported Alcohol
Involvement in the imputation process.
- Imputed BAC values are highly correlated with
Police-Reported Alcohol Involvement.
29th Annual Traffic Records Forum, Denver, CO 16
Comparison of Reported and Imputed values
for DRINKING=NOT Reported/UNKNOWN
People Saving People
National Center for Statistics & Analysis
- 74 percent of FINAL reported BAC values were 0.
- 70 percent of the imputed BAC values in the ARF
were 0.
- Even when there is no mention of probable cause, or
the lack thereof, for alcohol-testing, imputed BAC
values are in close agreement with the reported
BAC values.
29th Annual Traffic Records Forum, Denver, CO 17
Conclusions
People Saving People
National Center for Statistics & Analysis
States should not report alcohol involvement based only on
reported values
- Other cases may not have reported BAC values but may have
characteristics that point to a high likelihood of alcohol
involvement, or the lack thereof.
- Varying rates of reporting year-to-year may not allow valid
assessment of trends and progress.
Imputation levels the field of comparison
- Provides BAC values for all cases, so valid comparisons can
be made.
- Also provides statistical measures of uncertainty, etc.
29th Annual Traffic Records Forum, Denver, CO 18
Conclusions
People Saving People
National Center for Statistics & Analysis
Timeliness of Reporting to FARS
- Annual Report File (ARF) is used for Annual
Publications and Traffic Safety Facts and hence gets
maximum attention.
- Recommendation to FARS analysts to gather and
code BAC values in time for the ARF.
Imputation performs well
- Even in cases where there is no mention of probable
cause, or the lack thereof, for alcohol testing.
29th Annual Traffic Records Forum, Denver, CO 19
People Saving People
QUESTIONS?
National Center for Statistics & Analysis
Jim
Hedlund
29th Annual Traffic Records Forum, Denver, CO 20