SUMMARY
Strategies to Improve Reporting of Impaired and Distracted Driving in Motor Vehicle Crashes

The misreporting of impaired and distracted driving has a direct impact on every aspect of roadway safety, from vision statements, planning and programming, and safety analysis to the evaluation and implementation of safety solutions. Misreporting can affect the development of Strategic Highway Safety, Highway Safety, and Vision Zero plans, as well as safety allocations made in the Highway Safety Improvement Program. Misreporting can bias the resource allocation amounts and locations of behavioral safety improvements, affecting the identification of countermeasures for safety improvements and the ability to assess the safety impacts of programs, policies, and laws. If an agency systematically misreports impaired or distracted driving crashes, these crashes may be undervalued as a priority and underinvested in behavioral, legal, and infrastructure strategies and countermeasures, both at individual locations and systemwide.

The primary objectives of this report were to develop procedures for determining the extent of misreporting impaired and distracted driving crashes and to develop a methodology for improving the reporting of impaired and distracted driving in motor vehicle crashes. The term misreporting refers to a variety of issues, including underreporting, overreporting, errors in crash data recording, and misclassification. Guidance and methodologies developed in this research will help states and jurisdictions identify the misreporting of impaired and distracted driving and improve crash data collection and analysis.

The misreporting of impaired and distracted driving is multifaceted. Several factors contribute to misreporting impaired and distracted driving:

• Crash report limitations and variability: While the Model Minimum Uniform Crash Criteria provide states with guidance on data elements and attributes, this guidance is often modified to suit individual state needs.
• Impairment testing practices: In states that do not distinguish the nature of impairment, or where prosecutors choose not to apply enhanced penalties for drug involvement in addition to alcohol involvement, officers may not investigate drug impairment if alcohol results are positive.
• Officer training and department policies: In states where the crash report is treated as a legal document and admissible in court, officers may be trained to avoid entering any information that they cannot support through their own observation, witness statements, or direct evidence.

Two methodologies were developed that agencies can use to identify misreported alcohol-involved, drug-involved, and distracted driving crashes. The first method is text classification, which uses text-mining techniques to efficiently scan large amounts of crash narratives. A second method, direct linkage, uses the linkage of crash data to both hospital data and toxicology data.

Using text classification, agencies can identify the underreporting of alcohol-involved, drug-involved, and distracted driving crashes. The method was developed using 4 years of crash data from Wisconsin, and three case studies were conducted using 5 years of crash data from Connecticut, Kentucky, and Wisconsin. The presumption is that certain crash types are documented in police crash reports but not properly flagged, making impaired and distracted driving difficult to query. Text classification uses the narrative portion of crash reports to recover certain types of crashes that were not flagged properly. A NoisyOR method is used to determine the probability of a crash being a specific type by combining the probability scores of unigrams (i.e., single words) and bigrams (i.e., consecutive words) in the narrative. Text classification is effective, considering agencies only need access to the crash report (including the narrative portion of the form). However, text classification can only be used to identify underreporting: the model performance can be poor if the text data are noisy, an arbitrary threshold value is needed for classification, and careful interpretation of results is necessary and time consuming. Where officers are reluctant to include information on the crash report that they cannot support with direct observations, the narratives are unlikely to assist with identifying misreporting.

Through the direct linkage method, agencies can identify underreported alcohol-involved crashes by comparing flags in crash data, such as suspected alcohol use, with International Classification on Disease version 10 codes (ICD-10-CM) in the linked hospital records. Direct linkage can estimate actual crash counts and capture additional information not recorded in crash data. However, the accuracy of the linkage depends on the linking approach (e.g., probabilistic, using wildcards, versus deterministic, using exact matches of unique personal identifiers). Furthermore, overreporting cannot be determined because of the limitations introduced by alcohol exclusion laws and the low rates of alcohol and drug screening at medical facilities. Agencies will need to determine which ICD-10-CM codes they will use as indicators for alcohol, noting that these lists will have to be updated in the coming years as the United States transitions to ICD-11. Determining whether drug impairment is underreported can be challenging with this method because of the increased number of ICD codes that will be needed for different classes of drugs, impairment from multiple drugs, impairment from over-the-counter drugs, and impairment from negative reactions to prescribed drugs.

In the same way, comparing flags in the crash data to testing results from linked toxicology data, agencies can identify specific misreported alcohol-involved and drug-involved crashes. Comparing these two data sets can help agencies estimate actual crash counts and capture additional information not recorded in crash data. However, accuracy of the linkage depends on the linking approach (i.e., probabilistic versus deterministic). Laboratory stop testing policies can also affect the usefulness of flag comparison when analyzing drug-involved crashes. If the laboratoryʼs practice is to limit testing for the presence of drugs to only those cases when the tests for alcohol return a result below the per se limit, the absence of a drug test may not mean no drugs other than alcohol were present.

This research effort originally set out to develop methods to estimate misreported crashes. However, as a result of using the two developed methodologies, it was discovered that accuracy improved with a more descriptive analysis of crash results identified as alcohol involved or drug involved. The text classification method identifies crashes by using information the reporting officer included in the narrative. Through a manual review process, one can also identify language consistent with either alcohol- or drug-involved reported crashes. Unless the officer includes language relating to impairment in the narrative, the crash should not be considered as alcohol involved or drug involved. Similarly, linking to hospital data and toxicology testing results shows only whether alcohol or drugs were present in the driverʼs system at the time of the test or medical treatment. These data sources do not report impairment, so the misreported crashes can only be considered alcohol involved or drug involved.