Drug presence in driving deaths in Maryland: Comparing trends and prevalence in medical examiner and FARS data

Abstract

Accurate and reliable information on drug use by road users is essential to inform safety policy development but the reliability of national data has been questioned. There are two primary repositories of drug test information from fatal motor vehicle crashes in Maryland: (1) the Fatality Analysis Reporting System (FARS), which is a national crash database managed by the US Department of Transportation, and (2) the Maryland Medical Examiner (ME). In this study, we compared drug test information for people killed in crashes in Maryland between the FARS system and ME from 2006 – 2018. As ME records are the primary source for the FARS data from Maryland, these two data sets should be closely correlated. We used probabilistic linkage to match FARS and ME cases and compared matched cases by individual drug group. Matching was achieved on 83 % of cases (N = 4803 matched pairs). ME data consistently indicated higher overall incidence and trends in the presence of depressants, narcotics, and stimulants in crash deaths. Sensitivity analysis using both strict and relaxed matching criteria did not change this result. Road safety policy and prevention efforts for crashes involving drugs and alcohol require an accurate understanding of both long-term trends and year-to-year changes in drug prevalence. These findings demonstrate the potential value of using ME data as source of drug test information for crash deaths in the United States.

1. Introduction

Despite decades of effort, driver impairment remains a leading factor implicated in motor vehicle fatalities in the United States (National Center for Statistics and Analysis (NCSA), 2019b). Every year more than 10,000 individuals die on U.S. roads in crashes in which a driver was over 0.08 percent blood alcohol concentration, accounting for about one-third of motor vehicle crash deaths annually (National Center for Statistics and Analysis (NCSA), 2019b). Even as the alcohol-impaired driving problem continues, there is concern that the prevalence of additional potentially-impairing substances among drivers such as cannabis and opioids may be increasing. The prevalence of drivers testing positive for at least one drug in roadside surveys increased from 16 percent in 2007 to 20 percent in 2017 (Kelley-Baker et al., 2017). A large-scale National Highway Traffic Safety Administration (NHTSA) study which tested the prevalence of drugs and alcohol among drivers in Virginia Beach, Virginia found cannabis in 6.6 percent all participating drivers, including control and crash-involved drivers, the highest of any drug category (Lacey et al., 2016). However, while the prevalence of drug use may have gone up actual data on the extent of crash risk from drugs consumed by drivers is less clear.

To better understand crash risk, accurate data about the frequency of drug use by those involved in motor vehicle crashes are needed. While the Fatality Analysis Reporting System (FARS) provides the most comprehensive data across hundreds of variables on fatalities resulting from motor vehicle crashes in the United States, drug test data from toxicology tests reported in FARS have a number of limitations. Until 2018, the FARS database restricted the number of drugs that can be reported to three. This restriction meant that crashes involving polysubstance use could not comprehensively documented. There is also variation among and within states regarding which crash fatalities are tested, the types of drugs tested for, the type of specimen used (blood, oral fluid or urine), and details of test equipment and cut-off levels for confirming drug presence (Government Accountability Office (GAO), 2015; Berning and Smither, 2014). Recognizing the limitations of these data, NHTSA issued a Research Note advising against their use in studies of drug prevalence or trends (Berning and Smither, 2014). The agency also took steps to improve and expand drug test information in FARS, such as removing the three-drug reporting limit in 2018 (Berning and Smither, 2014; National Center for Statistics and Analysis (NCSA), 2019a; Slater et al., 2016). Nevertheless, considerable limitations remain and these continue to impede efforts to interpret FARS drug data.

Given the prominence of FARS data for research on crashes in the United States and the increasing attention to the potential impact of drug use on motor vehicle crashes, it is critical that we develop a better understanding of the accuracy of FARS drug test data. One approach to developing this understanding is to compare FARS drug test results to their primary source. Medical Examiners (ME) are designated by law to investigate the cause of deaths resulting from injuries, including motor vehicle crashes. MEs routinely sample blood, oral fluid, urine and other tissues as part of an autopsy of crash victims who died at the scene or those who died en-route to hospital (Slater et al., 2016). These samples are tested for the presence of drugs, although which substances are tested for, and at what thresholds, vary by state. Where toxicology data for crash deaths autopsied by the ME are available, these data are used as the source of FARS drug testing information (Jodon, 2019). Directly accessing the ME data may be an important and underutilized strategy for understanding the prevalence of alcohol and drugs among road users.

In this study, we compare FARS drug test data for Maryland to toxicology data for all crash deaths autopsied by the Maryland Medical Examiner spanning a 13-year period (2006–2018). All unexpected deaths occurring within Maryland are processed at a centralized facility in Baltimore, including toxicological testing of tissues and database management. This offers a high degree of consistency and quality control for these data. Toxicology data from the Maryland ME are the primary source of drug test information for Maryland crash deaths (Maryland Department of Transportation (MDOT), 2018) and have been previously analyzed to examine trends in opioid-involved crash deaths (Duren et al., 2019). Quantifying the agreement between FARS and ME toxicology data may provide further understanding of the strengths and limitations of FARS data.

2. Methods

2.1. Data sources

A description of the range and nature of the ME and FARS data used in this study follows:

2.1.1. Toxicology data from the medical examiner in Maryland

Toxicology data were obtained from the ME for drivers, passengers, motorcyclists, pedestrians and bicyclists killed in crashes. Under Maryland Statute §5—309, an autopsy is required for deaths resulting from injury. The majority of motor vehicle crash fatalities in Maryland are examined by the ME. In certain instances, an autopsy may not be performed if the death occurs while the person is hospitalized or if the family objects. Comparing the ME cases to reported numbers of overall crash fatalities in Maryland, ME cases represent 70–88 % of all crash deaths over the past decade, depending on the year.

Comprehensive toxicological testing is conducted as part of the autopsy. The majority of toxicology tests performed by the ME use blood samples (59 percent of all tests); the ME also uses urine (28 percent of tests), and other specimen such as vitreous, bile, and liver samples. The ME uses enzyme-linked immunosorbent assay (ELISA) for screening of morphine, benzodiazepines and oxymorphone.

Gas chromatography and mass spectrometry (GCMS) for screening and confirmation for all other substances. No minimum cutoff is used for the concentration of any substances. Except in rare cases, the ME in Maryland does not test for cannabis; therefore, cannabinoids are excluded as a drug category from these analyses.

The toxicology data are stored in an electronic database that can be queried by cause of death and year. The investigators requested and received data from the ME database for all motor vehicle crash deaths from January 1st 2006 to December 31st 2018, which included 6131 unique cases. Deaths that involved non-roadway vehicles were excluded, such as boats, aircraft, and trains; which eliminated 178 cases. Intentional deaths (as determined by the medical examiner) were also excluded, including homicide and suicide cases; this eliminated an additional 165 cases. An additional 58 cases that fell outside of the study period were eliminated. After applying these exclusions, the total cases decreased from 6131 to 5,768.

2.1.2. FARS data

FARS data are managed by NHTSA. FARS data were downloaded for each year of analysis (2006–2018) and the person file, accident file, vehicle file, and drug files were merged. For the purpose of this analysis, the data were limited to fatal crashes occurring in Maryland spanning January 1st 2006 to December 31st 2018. At the time these analyses were conducted, the final 2018 FARS data were not available, therefore data from the Annual Report File data from NHTSA were used (National Highway Traffic Safety Administration (NHTSA), 2020). This resulted in 6898 cases from FARS, which included drivers and other road users, including pedestrians and bicyclists.

2.2. Data definitions

2.2.1. Positive drug tests

All of the substances listed in the Maryland ME data were cross-referenced to substances listed in the FARS Coding Manual (National Highway Traffic Safety Administration (NHTSA), 2018). Any substances in the ME data not found in the FARS coding manual were individually classified according to the drug categories in the manual. This manual classification was done in collaboration with the University of Maryland School of Pharmacy’s Maryland Poison Center. For both data sources, a positive case is defined as a death where the presence of any substance was greater than zero.

2.2.2. Year

Year corresponds to the year of the crash; a definition that is consistent with FARS classification. Across all years of data, there were 466 missing crash dates in the ME data; for these cases, the year from the date of death variable was used to impute the year of the crash. There were no missing data for year of crash in FARS across the study period.

2.2.3. IRB approval

This study was reviewed by the Johns Hopkins Institutional Review Board who determined that fatally-injured crash victims do not qualify as human subjects research under DHHS regulations 45 CFR 46.102, and the study does not require IRB oversight.

2.3. Analysis

We conducted a probabilistic linkage to match individual cases in the FARS and ME data sets and compared drug presence in the matched cases. Eight variables relating to crash and demographic information were present in both datasets; including age, sex, race, road user type (driver, passenger, motorcyclist, bicyclist, pedestrian), crash city and county, the crash date and the date of death, and these were used as the matching criterion. An expectation-maximization algorithm was then used to generate probabilities of agreement for each possible pairing of observations based on the variables contained in both datasets. Of these results, the pairing with the highest probability of agreement was selected, limiting the matched pairs to a one-to-one match for each FARS and ME matched case. The results were further constrained to only include those that matched on either the crash date or date of death as well as an additional four variables. The pairs identified by this probabilistic matching method were used to compare toxicology data between FARS and the ME data. For this analysis we did not examine differences between road user types, as the primary purpose was to compare toxicology reporting and the accuracy of such reporting should not vary among road user types.

Positive drug test information from FARS and ME data were compared in matched cases, using two proportion Z-tests and Welch Two Sample t-tests. Trends in matched toxicology data were also assessed over time using the Cochran-Armitage test. The statistical software R 3.6.3 and the tidyverse package were used for all of the data analysis and related figures completed as part of this work (R Core Team, 2020; Wickham et al., 2019).

3. Results

3.1. Prevalence of drug-positive cases in FARS and ME data

A total of N = 6898 FARS cases and N = 5768 ME cases met the study inclusion criteria. Across the study period (2006–2018), the Maryland ME tested 100 % of cases they investigated for alcohol and 94 % for other drugs. It was not possible to estimate the prevalence of testing in FARS data because while positive test results are reported, those that test negative are not reported. Probabilistic matching resulted in a total of 4803 matched pairs between the FARS and ME data for the years 2006–2018. This corresponds to 83 percent of all ME cases (5768 total cases) and 70 percent of all FARS data (6898 total cases), as shown in Fig. 1.

Fig. 1.

Venn diagram of FARS, ME and matched cases (2006–2018).

Table 1 presents the number of drug-positive cases among the matched cases by the data source and by substance type across the study period. Agreement on drug presence between FARS and ME data ranges from 85 to 100 percent, although the highest agreement occurs for substances with zero or near-zero positive cases in both datasets. The percentage of cases with hallucinogens, depressants, stimulants, narcotics, and other drugs present were significantly higher in the ME data relative to FARS.

Table 1.

Agreement between positive drug tests for matched FARS and ME data (2006–2018).

Substance Type	Percent Agreement [%]	FARS Positive N [%]	Medical Examiners Positive N [%]
Anabolic Steroid	100.0	0 [0.0]	0 [0]
Inhalant	99.9	0 [0.0]	3 [0.1]
PCP	99.5	76 [1.6]	82 1.7]
Hallucinogen	98.6	26 [0.5]	83 [1.7]**
Depressant	98.1	114 [2.4]	172 [3.6]
Stimulant	96.5	320 [6.6]	426 [8.9]**
Alcohol	95.5	1741 [36.2]	1711 [35.6]
Narcotic	95.4	361 [7.5]	474 [9.9]**
Other Drug	85.2	688 [14.3]	1020 [21.2]**

^**ME drug test prevalence significantly higher than FARS (p < .001 level).

To determine the extent to which the FARS data were a subset of the ME data, we examined the overlap in drug-positive results between the two sources. In three out of four cases (75 %), substances in FARS and ME were identical. In approximately one in five cases (19 %) there was at least one substance found in the ME data that was not present in FARS data.

To investigate the source of the greater number of drug-positive results in ME data, we considered the role of the three-drug reporting limit in FARS prior to 2018 (Table 2). As a result of this limitation, it could be expected that the ME data may include a greater number positive drug results, as the ME data have no limit on the number of drug test results. Among the 4803 matched cases, 4443 (92.5 %) occurred prior to 2018 when the FARS limitation was in place. Of these 4443 pre-2018 cases, 4238 (95 %) had no third substance indicated in FARS. The remaining 205 cases (5%) had three substances listed in FARS, and of these cases, 115 (56 percent) had more than 3 drugs listed in the ME data. Given this information, the limitation of only recording three drugs in FARS could only have impacted the toxicology results in 115 of the total 4803 matched cases (2.4 percent). From 2018 onwards, when there is no limit on drug test results in FARS, none of the matched cases from FARS had more than three positive drug tests listed.

Table 2.

Analysis of differences between FARS and ME matched cases for identified drugs.

Description	Number of cases
Total cases	4803
Pre-2018 cases	4443
Pre-2018 cases ≤2 drugs indicated in FARS	4238
Pre-2018 cases with 3 drugs indicated in FARS	205
Pre-2018 cases with 3 drugs indicated in FARS and more than 3 drugs indicated by ME	115

To investigate an additional potential discontinuity between the FARS and ME matched cases, we analyzed the extent to which matched FARS cases identified a positive substance that was not found in the ME data. In 387 or 8% of matched cases, FARS had at least one positive substance that the ME did not. Table 3 lists the positive drug tests in FARS cases that were not identified in the ME data. The majority of these (79 %) were positive for alcohol or the Other Drug category in FARS but not in the ME data. In 40 of the 387 cases, the FARS data had more than two substance categories identified as positive in FARS but not positive in ME. These test results in FARS come from a source other than the ME data and since that source is unknown, no judgment can be made from this study as to the accuracy of the reporting.

Table 3.

Frequency of cases with positive FARS results not found in ME data^*.

Substance Type	Number of cases (N = 387)
Anabolic Steroid	0
Inhalant	0
PCP	9
Hallucinogen	4
Depressant	17
Stimulant	33
Alcohol	124
Narcotic	54
Other Drug	189

^*Categories are not mutually exclusive.

Fig. 2 presents trends for the matched FARS and ME positive drug tests for alcohol, depressants, narcotics, and stimulants. The ME data had a higher prevalence of drug-positive cases across all years between 2006 and 2018 for all substances except alcohol. Moreover in 2018, the ME data suggested a greater increase across all categories of substances, with narcotics, for example, increasing in the ME data from 13.4 % in 2017 to 16.4 percent in 2018 while increasing from 11 % in 2017 to 13 % in 2018 in the FARS data. Both FARS and ME matched datasets show statistically significant increases in the rates of positive cases for depressants and narcotics (Z=−3.25 and −4.81 for ME data, respectively, and Z=−425 and −8.76, respectively, for the FARS data; p-values<.001) but show no significant trend in the rates of stimulant-positive cases. Notably, the two data sets also show drug-positive rates moving in different directions in several instances. For example in 2015, the ME data show increases in depressants and stimulants while the FARS data show decreases.

Fig. 2.

Comparison of trends in positive drug tests among matched FARS and ME cases (2006–2018).

3.2. Sensitivity analysis

To assess the sensitivity of the results to the approach used in the probabilistic linking, additional matches were run with more- and less-restrictive constraints. Whereas the data linkage model used in the study analyses required matching at least 5 variables between FARS and ME cases and resulted in 4803 matched pairs, a conservative model was created that forced a match on at least 7 of the 8 linkage variables and resulted in 2648 matched pairs. Additionally, a less restrictive model was created that matched on either the crash date or date of death and at least two additional linking variables, resulting in 5052 matched pairs. The same trends were observed in these two models as were seen in the original model depicted in Fig. 2. The difference in prevalence statistics between FARS and ME data as well as the prevalence values in the sensitivity analysis remained relatively constant; the prevalence values never differed by more than one percent from the primary match set. For instance, the prevalence for narcotic cases among the conservative matched data was 9.9 percent in the ME data and 7.9 percent in FARS. For the less stringent matching, the narcotics prevalence was 9.9 percent in the ME data and 7.2 percent in the FARS data. This sensitivity analysis demonstrates that the main study findings are not strongly reliant on the specific matching constraints, but rather seem generalizable to both more conservative as well as more relaxed matching criteria.

The analysis presented in the preceding section relies on the assumption of the ME data as the sole source of toxicology data for FARS analysts in Maryland where a motor vehicle crash victim did not die in hospital. While this may be the case, it may also be true that these analysts supplemented ME toxicology data with information from other sources, such as police reports. If other sources were used, this could help explain the 8 percent of cases identified earlier as having a positive substance identified in FARS that was not found in the ME data. To reduce our reliance on the assumption of the ME data as the sole source of toxicology information, we repeated our reported analysis from the preceding section without the 387 cases that had positive FARS drug classifications that were absent in the ME data. Without this subset of matched cases, the prevalence for each drug category remained relatively constant for the ME matched cases as compared to our earlier analysis, always differing by less than 1 percent, if at all. In contrast, the prevalence in the FARS matched cases were lower across all drug categories (except for anabolic steroids, which has no positive matched FARS cases). For example, the narcotics prevalence decreased from 7.5 percent to 5.9 percent and the Other Drug category prevalence decreased from 14.3–10.4 percent. The trend analysis presented the same conclusions as before, although in some years the FARS matched data showed lower prevalence rates. Thus, the overall conclusions remain the same, although the prevalence differences between ME and FARS data became more pronounced in this additional analysis.

4. Discussion

To be useful for informing or evaluating drugged driving policies, data on drug use by drivers needs to be accurate and reliable. This study was conducted to better understand the strengths and limitations of drug test data from the FARS system, by comparing information reported for fatalities in Maryland with data for the same individuals from the ME, which is the source for the FARS data. Similar to previous studies comparing FARS data to source data, (Bunn et al., 2019, 2013) we found generally moderate levels of agreement between FARS and the ME data.

While it was not possible to obtain a high probability match of all individuals in the two data sets, our methodology allowed 76 percent of total cases (83 percent of all ME cases and 70 percent of all FARS data) to be paired. Differences in inclusion criteria for crash deaths, and structural differences in the data sources may explain the inability to match the remaining cases. For example, to be included in FARS, a crash death must involve a motor-vehicle on a public trafficway and must occur within 30 days of the crash (National Center for Statistics and Analysis (NCSA), 2019a). By contrast, the ME considers crash deaths occurring on private roads, as well as cases that extend beyond 30 days. Our probabilistic matching method used individual and crash-related factors but excluded key identifiers such as name or date of birth. Previous studies using probabilistic matching to match police crash reports to hospital, emergency department, or EMS data, ranged from 75 to 100 percent of cases (Milani et al., 2015, 26). These studies typically used first name, last name, and date of birth (Cook et al., 2015, 35). As these identifying variables were not available in FARS, the match rate achieved in this study is comparable to previous studies. Further investigation into potential causes for the failure to match 17 percent of cases may allow higher matching rates in future studies. However, a sensitivity analysis using more- and less strict matching requirements indicates that refining the match rate for this study would not alter the overall findings.

The ME data had a higher number of drug-positive deaths relative to FARS in both the percentage of cases and absolute numbers. While the percentages allow comparisons across the substance types, the absolute numbers are informative to understand rare events. For example, the discordance between FARS and ME data for hallucinogens across the study period was over three fold. FARS included 26 deaths where hallucinogens were involved, corresponding to 0.5 % of deaths during the study period. In contrast, the ME included 83 hallucinogen-involved deaths corresponding to 1.7 % of cases.

Structural differences may have contributed to variation in drug test reporting between the two data sets. Prior to 2018, FARS reporting was limited to a maximum of three drug substances per case (National Center for Statistics and Analysis (NCSA), 2019a) whereas the ME has no limit on the number of substances. Thus, to the extent that the ME recorded more than three drugs, this difference could contribute to a higher prevalence of positive drug results observed in the ME data. However, an analysis of the number of cases during the study period where the ME data included more than three drugs for an individual indicates that this 3-drug limitation in the FARS data likely had a very limited effect on agreement between the two data sets.

Since the ME data are the primary source of FARS reporting on driver toxicology in Maryland, accessing this information directly from the source might avoid some of the misalignment in the FARS data observed in this study. However, other structural differences may affect the potential representativeness of ME data. As noted previously, the ME does not perform toxicological tests on every Maryland traffic death and this incomplete sample could affect the degree to which the ME findings reflect drug use among all those killed in crashes. Over the years covered in this study, the ME conducted drug tests on between 70 and 88 percent of Maryland crash deaths, the remainder consisting primarily of deaths that occurred under the care of a physician who could determine the cause of death without the need for an autopsy. To investigate the representativeness of the ME data, a prospective comparison of ME cases to non-autopsied crash deaths is needed.

5. Limitations

It is important to note that these analyses are based on drug-positive cases only. A comparison of negative drug tests between the FARS and ME data was beyond the scope of this study. The positive drug results are accurate when aggregating the specific drugs to their corresponding FARS drug category, however, there may still be differences among these cases for the specific drugs identified or the number of specific positive drugs identified that fall within the same drug category. At the time these analyses were conducted, the final 2018 FARS data were not available, therefore data from the Annual Report File data from NHTSA were used (NHTSA 2020). Although the 2018 data file is a full year’s worth of data, it may be subject to change when it is finalized.

6. Conclusions

Direct comparison of the incidence and trends in substance-involved crash deaths between FARS and Maryland ME data provides new insights into the advantage of using original source data to obtain a high-resolution understanding of the presence of impairing substances in crash deaths. Since the ME records are the primary source for the FARS data, we had expected these two data sets to be closely correlated. However, ME data consistently indicated higher overall drug incidence, showed differences in trends in drug-positive crash deaths, and in the case of some drugs, showed large proportional differences in counts of drug-positive cases. While overall trends between the FARS and ME data were generally aligned, pronounced year-to-year differences were evident, suggesting systematic underreporting in FARS. The differences observed between the two data sets mean that conclusions drawn from the FARS toxicology data with regard to issues such as evaluation of a drug-impaired driving law or changes in drug prescribing practices may lead to different policy recommendations than would have been indicated by the source ME data.

The observed misalignment between FARS and the source ME data suggest a problem with incomplete reporting to FARS and the need for more accurate and reliable toxicology data to inform policy decisions. Such data improvements could be achieved through multiple pathways, including improving the accuracy of the FARS system or through the creation of a new system. The robust data collection protocols of the ME provide stable and reliable toxicology data for crash deaths. Currently, sixteen states have medical examiner systems and most of these are statewide (Centers for Disease Control and Prevention, 2020). Some state systems are centralized and others use regional facilities. Prior to inclusion in any new system, there is a need to evaluate medical examiner policies with regard to autopsies and toxicology testing of crash victims.

The findings of this study demonstrate the potential value of ME data as a standalone source of drug test information for crash deaths in the United States. Future work may consider the development of a sentinel surveillance system based on a selected sample of rigorous MEs across the nation that provide high-quality and reliable drug testing information for motor vehicle crash deaths (Kelley-Baker et al., 2019).