I-95 high-risk driver analysis using multiple imputation methods

Size: px

Start display at page:

Download "I-95 high-risk driver analysis using multiple imputation methods"

Neil Joseph
5 years ago
Views:

1 I-95 high-risk driver analysis using multiple imputation methods Kyla Marcoux Traffic Injury Research Foundation New Orleans, Louisiana July 26, 2010

2 Acknowledgements Authors: Robertson, R., Wood, K., Vanlaar, W. and Simpson, H. I-95 Corridor Coalition 2

3 Overview Background Methods Results Discussion Conclusions 3

4 Background The I-95 Corridor Coalition is a group of transportation agencies and related organizations across sixteen member States and the District of Columbia along the I-95 corridor with a goal of improving transportation system performance. 4

5 Background High-risk drivers (HRDs) are often described as a relatively small group of persistent traffic violators (usually less than 10%) It is believed that these persistent offenders are responsible for a significant portion of the serious injury and fatal collisions on the highways The high risk driver problem has not, until recently, received much attention Data on the dimensions of the problem are limited 5

6 Background Purpose of the Ten Percenter project: To determine who the HRDs are; and To identify the magnitude and characteristics of the HRD problem Goal is to address the problem of the ten percenters or HRDs to improve the safety for the traveling public. 6

7 Method Operational definition of HRDs A driver who had a BAC of 0.16% or above at the time of the crash, OR refused a breath test, OR was involved in 3 distinct events in the last 3 year. These events include: impaired driving offences speed violations other types of moving violations previously recorded crash(es); or, license suspensions 7

8 Method Magnitude and characteristics of HRDs FARS analyses FARS analyses with multiple imputation data Difficulty with first analysis Much of the BAC data are missing in FARS Drivers could be identified as high-risk using all the agreed to criteria except BAC 8

9 Method To overcome this problem Second analysis created a BAC value where it was missing using a wellestablished procedure Rubin s (1979) multiple imputation 9

10 Method Multiple imputation NHTSA relies upon Rubin s multiple imputation process to create sound statistical estimates for the missing BAC values NHTSA uses a variety of characteristics including police-reported drinking, age, gender, restraint use, type of crash, time of day, and driver of the striking or struck vehicle to determine a distribution of alcohol consumption for each missing data point. 10

Method Multiple imputation The distribution of possible BAC values (10 for each missing BAC value) is then used in the analyses As if the analyses are conducted ten times, each time replacing the

11 Method Multiple imputation The distribution of possible BAC values (10 for each missing BAC value) is then used in the analyses As if the analyses are conducted ten times, each time replacing the missing BAC value with a replacement value coming from the distribution of possible BAC values This multiple imputation method summarizes the results for each of the ten data runs into one single result 11

12 Crash analysis using FARS In total, approximately 14% of drivers involved in fatal collisions were considered to be HRDs. The percentage ranged from a low of 3% in DC to a high of 19% in NJ. 12

13 Characteristics of HRDs HRDs were more commonly involved in single vehicle collisions where the vehicle ran off the road and hit a fixed object Drivers in these collisions tended to be male, aged 21-34, unbelted, speeding, under the influence of alcohol or drugs, and were likely to have an invalid license Collisions most often occurred on weekends, at night, and when it was dark HRDs represent a small proportion of drivers but account for a very substantial portion of fatal injury collisions 13

14 Multiple imputation crash analyses using FARS In total, approximately 25% of drivers involved in fatal collisions were considered to be HRDs. The percentage ranged from a low of 15% in DC to a high of 33% in CT. This average (25%) is now relatively close to findings from other research about the hard core drinking driver that have shown that almost one third (27%) of all fatally injured drivers (in 1995) were hard core drinking drivers (BACs of 0.15% or above). 14

15 Multiple imputation crash analyses using FARS For each jurisdiction The percentage in the Don t know category is lower for the results using MI data compared to those without The percentage of HRDs is consistently and considerably higher The percentage of non-hrds is lower in all jurisdictions (with the exception of GA which is 2% higher) when using MI data compared to the results without MI data. 15

16 Multiple imputation crash analyses using FARS The differences between HRDs and non- HRDs in terms of magnitude and characteristics are more pronounced when using multiple imputation data. However, there a few cases where the differences were the same or smaller not more that 3% smaller 16

17 Discussion This issue of missing BAC data (over 60%) would not have allowed the use of the operational definition of the study Multiple imputation is a robust statistical technique The conditions of use for MI data were met (large sample) 17

18 Conclusions As expected, this inclusion of BAC data effectively increased the number of high-risk drivers It also allowed for the use of more advanced statistical analyses which are unstable when there are too many missing values 18

19 Conclusions Overall, multiple imputation served as a valuable tool in these analyses by allowing the use of more complete data This allowed the research objectives of this study to be fulfilled and the definitions of the study to be applied to the analyses 19

20 Staying informed 20

I-95 Corridor-wide safety data analysis and identification of existing successful safety programs. Traffic Injury Research Foundation April 22, 2010

I-95 Corridor-wide safety data analysis and identification of existing successful safety programs Traffic Injury Research Foundation April 22, 2010 Overview Background Methodology Purpose Crash analysis