research paper/ Findings 3 pages within 12hrs please

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

The effects of medical marijuana laws on cannabis-involved driving

Eric L. Sevigny Georgia State University. Andrew Young School of Policy Studies, Department of Criminal Justice and Criminology, 55 Park Place NE, Suite 519, Atlanta, GA 30303, USA

A R T I C L E I N F O

Keywords: Medical marijuana laws Cannabis-involved driving Differences-in-differences Multiple imputation Fatality analysis reporting system

A B S T R A C T

This study uses data from the Fatality Analysis Reporting System and a differences-in-differences model to examine the effect of state medical marijuana laws (MMLs) on cannabis-involved driving among U.S. drivers involved in a fatal crash between 1993–2014. Findings indicate that MMLs in general have a null effect on cannabis-positive driving, as do state laws with specific supply provisions including home cultivation and un- licensed or quasi-legal dispensaries. Only in jurisdictions with state-licensed medical marijuana dispensaries did the odds of marijuana-involved driving increase significantly by 14 percent, translating into an additional 87 to 113 drivers testing positive for marijuana per year. Sensitivity analyses reveal these findings to be generally robust to alternate specifications, although an observed spillover effect consistent with elevated drugged driving enforcement in bordering states weakens a causal interpretation. Still, reasonable policy implications are drawn regarding drugged driving prevention/enforcement and regulations governing dispensary delivery services and business siting decisions.

1. Introduction

The growing prevalence of marijuana-involved driving in the U.S., along with increasingly robust evidence that cannabis-impaired driving significantly increases traffic crash risk, has heightened public concerns over roadway safety (Rogeberg and Elvik, 2016). These concerns are particularly salient in light of the ongoing liberalization of state mar- ijuana laws allowing legal access to cannabis for medical and recrea- tional purposes (Hall, 2015; Huestis, 2015; Pacula and Sevigny, 2014). As of year-end 2016, twenty-eight states and the District of Columbia have passed medical marijuana laws granting authorized patients the right to use marijuana therapeutically, and eight states plus the District of Columbia have outright legalized adult recreational marijuana use. With states continuing to enact and amend these laws at a rapid pace, policymakers require reliable evidence of their impact on marijuana- involved driving and roadway safety. Despite intense public interest, however, evidence in this policy space remains limited, as only a handful of empirical studies have directly investigated this association. To help fill this gap, this paper uses individual-level data from the Fatality Analysis Reporting System (FARS) and a differences-in-differ- ences specification to examine the effect of state medical marijuana laws (MMLs) on marijuana-involved driving among all drivers involved in a fatal crash in the U.S. between 1993–2014.

1.1. Medical marijuana laws, marijuana-involved driving, and roadway safety

Marijuana is the most commonly detected illicit substance among drivers (Berning et al., 2015; Brady and Li (2014); Romano and Pollini, 2013; Rudisill et al. 2014). Toxicological evidence from the National Roadside Survey indicates that marijuana positivity rates jumped from 8.6% to 12.6% between 2007 and 2013–2014 among weekend night- time drivers aged 16 or older (Berning et al., 2015), and the Fatality Analysis Reporting System reveals significant longer-term increases in marijuana use detected in drivers involved in fatal vehicle accidents, increasing from 29% to 37% between 1993 and 2010 among drug-po- sitive drivers (Wilson et al., 2014). This evidence raises substantial public safety concerns, particularly against the backdrop of marijuana liberalization laws that increase legal access to cannabis, as cannabis- impaired driving has been shown to significantly increase individual traffic crash risk by roughly 20–30% (Rogeberg and Elvik, 2016).

Despite the apparent links between medical marijuana laws, im- paired driving, and crash risk, just a handful empirical studies have investigated the effects of MMLs on roadway safety. Three studies use data from the National Highway Traffic Safety Administration’s (NHTSA) Fatality Analysis Reporting System (FARS) to examine can- nabis-positive driving among drivers involved in fatal vehicle crashes. Masten and Guenzburger (2014) analyzed changes in cannabis-positive driving for the period 1992–2009 using a series of interrupted time series analyses for twelve MML states, finding significant increases in

https://doi.org/10.1016/j.aap.2018.05.023 Received 4 January 2018; Received in revised form 3 April 2018; Accepted 31 May 2018

E-mail address: [email protected].

Accident Analysis and Prevention 118 (2018) 57–65

Available online 07 June 2018 0001-4575/ © 2018 Elsevier Ltd. All rights reserved.

T

cannabis-positive driving in just three states (California, Hawaii, and Washington). Similarly, Salomonsen-Sautel et al. (2014) employed time series methods to compare changes in driver marijuana positivity be- tween 1994–2011 in Colorado relative to aggregate toxicological data on drivers from 34 non-MML states. Their analysis, which used the 2009 commercial expansion of the Colorado medical marijuana in- dustry as the analytic break point rather than the state’s MML enact- ment year of 2001, found Colorado driver positivity rates increased significantly post-commercialization relative to non-MML states. Most recently, Hamzeie et al. (2017) used a driver-level panel dataset for the years 2010–2014 to investigate the effects of MMLs on cannabis posi- tivity among drivers, controlling for driver, crash, and vehicle char- acteristics. Unfortunately, because the authors employed a stepwise regression procedure in which the MML indicator was dropped from the model due to nonsignificance, we can only conclude there was a null policy effect.

Two other studies of MMLs also use FARS data, but investigate changes in traffic fatality rates. Performing a differences-in-differences analysis of state-level data for the years 1990–2010, Anderson et al. (2013), found no statistical evidence that MMLs affect overall crash risk after controlling for state-level demographics, competing state policies, and state-specific time trends. However, supplemental analyses showed MMLs were associated with statistically significant declines in alcohol- involved fatal accidents, which the authors interpreted as evidence that MMLs reduce crash risk through the individual-level mechanism of marijuana-for-alcohol substitution. Lastly, Santaella-Tenorio et al. (2017) investigated the effects of MMLs on the traffic fatality rate for the period 1985–2014, finding an 11% reduction in the traffic fatality rate post-MML implementation, net of controls for rival state policies, annual highway safety expenditures, vehicle miles travelled, and al- cohol consumption. Additional analyses focusing on the impact of dis- pensaries revealed null effects on the traffic fatality rate, suggesting that this supply mechanism has no direct effect on aggregate traffic crash rates.

Overall, these five studies provide inconsistent evidence on the impact of MMLs on marijuana-involved driving and associated crash risk. On the one hand, MMLs appear to have null effects on the pre- valence of cannabis-positive driving, although some states with more commercialized medical marijuana industries seem to have experienced significant increases in driving proximal to marijuana use. Conversely, studies examining traffic fatalities suggest that MMLs do not increase crash risk and, in fact, may lower the traffic fatality rate through the mechanism of alcohol-to-marijuana substitution.

1.2. Current study

Using driver-level data from the Fatality Analysis Reporting System (FARS), I perform a differences-in-differences analysis within a multiple imputation framework to investigate the effect of state medical mar- ijuana laws on marijuana-involved driving among all U.S. drivers in- volved in a fatal vehicle crash between 1993–2014. The present study extends prior research in several ways. First, the use of driver-level FARS data provides a more granular level of control over individual and contextual factors that may confound the association between MMLs and driving behavior. Among prior studies, only Hamzeie et al. (2017) analyzed microlevel data, but they focused on a relatively limited period (i.e., 2010–2014) that did not completely encompass pre-policy years. Second, the study employs multiple imputation to address missing FARS data; no prior studies have addressed missing data issues with a similar degree of rigor. Third, the study investigates hetero- geneous effects of MMLs by examining the impact of legal provisions governing the supply of medical marijuana (i.e., cultivation, dis- pensaries). Among prior studies, only Santaella-Tenorio et al. (2017) investigated the policy effects of dispensaries (but not personal culti- vation).

2. Methods

2.1. Data and measures

FARS data represent a census of all motor vehicle crashes on public roadways that result in a motorist or non-motorist fatality within 30 days of the crash. Information is coded on more than 100 data elements for each fatal accident from a variety of sources, including police re- ports, death certificates and medical examiner reports, vehicle regis- tration and driver licensing files, emergency medical service and hos- pital records, and highway department data (National Highway Traffic Safety Administration, 2015). For this study, I merged annual FARS data files at the driver level of analysis for the years 1993–2014, ex- cluding operators less than 12 years old, resulting in an analytic file of nearly 1.2 million drivers who were involved in a fatal vehicle accident over the 22-year study period. Relevant driver and contextual factors were then coded for each case, carefully accounting for annual changes in FARS coding practices (National Highway Traffic Safety Administration, 2015). Full coding syntax documenting oper- ationalization of measures is provided in the online supplement.

The primary dependent variable, cannabis+, measures whether the driver tested positive for THC or related metabolite on up to three re- corded drug tests. Policy variables were operationalized based on a comprehensive review of state statutes, regulations, and relevant court cases; legal summaries of state laws; and published reports (e.g., Governor’s Highway Safety Association, 2015; Lacey et al., 2010). Given that specific crash dates are reported in FARS, policy indicators were “turned on” as of the effective date of these laws. Table 1 displays the effective dates used to code these policy variables as of December 31, 2014. Coding of the focal medical marijuana laws was based on the author’s review of state statutes and regulations, supplemented by ad- ditional secondary sources (Klieger et al., 2017; Marijuana Policy Project, 2016). Medical marijuana codes state laws that extend legal protections to medical marijuana patients (whether through exemption from arrest/prosecution or an affirmative defense).1 Additionally, home cultivation codes the effective date of provisions allowing patients to cultivate their own marijuana, whether universally or by special permit,2 and dispensaries identifies states with either (i) legally per- mitted and regulated medical marijuana dispensaries or (ii) illegal or quasi-legal medical dispensaries operating without licensure from the state.3

For the rival policy variables, marijuana decriminalization indicates state laws that eliminate jail time for the simple possession of small amounts of marijuana.4 Recreational marijuana identifies states that

1 In most states, this coincides with the law’s nominal effective data, but in four states (i.e., DC, IL, NH, VT) legal protections commenced by statute or rule only when patients were in possession of a registry card; for these states, the effective date is set when the state began issuing patient ID cards.

2 Most states have universal patient cultivation allowances. Arizona did until licensed dispensaries opened, whereupon patients were permitted to grow only if they resided more than 25 miles from a dispensary (which essentially eliminated home cultivation since most patients fell within this radius). New Mexico requires patients to obtain a personal production license (PPL). Two other states (MA, NV) allow patient cultivation only with a hardship exemption, but these provisions were not operational during the study period. Finally, although universal home cultivation was implicitly authorized in its original legislation, Washington did not explicitly authorize patient cultivation until a decade after initial law passage.

3 The effective date is coded according to when dispensaries first opened for business in order to capture actual supply effects. For legal dispensaries, identifying this date was relatively straightforward because state oversight agencies often report openings for both patient and general public knowledge, and the associated media coverage tends to be intense. For illegal and quasi-legal dispensaries seeking to avoid law enforcement at- tention, we had to rely on less reliable media accounts, law enforcement press releases, and other information to determine the operational dates of these storefront dispensaries. Additional state-specific historical or documentary research would improve our under- standing of these developments.

4 Note that California’s new decriminalization law, which went into effect on January 1, 2011, further reduced the penalty for possessing less than an ounce of marijuana from a

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

58

have legalized the possession and sale of marijuana for adults, which includes Colorado and Washington within the timeframe of our study. THC per se law indicates states that have enacted a universal impaired driving statute making it illegal for anyone to drive with a blood THC concentration above a specified limit.5

Additional measures were developed to account for rival state po- licies and enforcement practices relevant to roadway safety but not specific to marijuana. This includes the BAC 0.08 law, which is a per se law that makes it an impaired driving offense to operate a vehicle with a blood alcohol concentration of 0.08 or above (National Highway Traffic Safety Administration, 1998; National Institute on Alcohol

Abuse and Alcoholism, 2015). The intensity of roadway safety en- forcement is measured by the number of sworn police officers per 100,000 population (Federal Bureau of Investigation, 1994-2015)6 and number of - (DREs) per 100,000 population (DREs per 100,000) (International Association of Chiefs of Police, 1999-2015).7

A range of driver-level factors are employed as controls. Given that marijuana use is associated with alcohol and other drug use in drivers (Romano and Pollini, 2013; Scherer et al., 2013), driver positivity

Table 1 Effective dates of state marijuana laws as of December 31, 2014.

Statea Medical marijuana Marijuana decriminal-izationb Recreational marijuana THC per se law

MMLc Home cultivationd Dispensariese

Alaska 3/4/99 3/4/99 (U) – 1975 – – Arizona 12/10/10 12/10/10 (U)12/6/12 (P) 12/6/12 (L) – – 7/28/90 California 11/6/96 11/6/96 (U) 11/6/96 (O)1/1/04 (L) 1976 – – Colorado 6/1/01 6/1/01 (U) 1/1/05 (O)1/1/11 (L) 1975 12/10/12 5/28/13 Connecticut 10/1/12 – 8/20/14 (L) 7/1/11 – – Delaware 6/1/11 – – – – 7/10/07 District of Columbia 6/15/13 – 7/30/13 (L) 7/17/14 – – Georgia – – – – – 7/1/01 Hawaii 12/28/00 12/28/00 (U) – – – – Illinois – – – – – 8/15/97 Indiana – – – – – 7/1/01 Iowa – – – – – 7/1/98 Maine 12/23/99 12/23/99 (U) 4/1/11 (L) 1976 – – Maryland 10/1/03 – – 10/1/14 – – Massachusetts 1/1/13 1/1/13 (U) – 1/1/09 – – Michigan 12/4/08 12/4/08 (U) 12/1/09 (O) – – 9/30/03 Minnesota 5/30/14 – – 1976 – – Montana 11/2/04 11/2/04 (U) 12/1/09 (O) – – 10/1/13 Mississippi – – – 1977 – – Missouri – – – – – – Nebraska – – – 1979 – – Nevada 10/1/01 10/1/01 (U) 3/1/10 to 7/1/11 (O) 10/1/01 – 9/23/03 New Hampshire – – – – – – New Jersey 10/1/10 – 12/6/12 (L) – – – New Mexico 7/1/07 4/15/08 (U)12/15/08 (P) 7/31/09 (L) – – – New York 7/5/14 – – 1977 – – North Carolina – – – 1977 – – Ohio – – – 1975 – 8/17/06 Oklahoma – – – – – 10/1/13 Oregon 12/3/98 12/3/98 (U) 11/1/09 (O)3/3/14 (L) 1973 – – Pennsylvania – – – – – 2/1/04 Rhode Island 1/3/06 1/3/06 (U) 4/19/13 (L) 4/1/13 – 7/1/06 Utah – – – – – 5/2/94 Vermont 10/28/04 7/1/04 (U) 7/18/13 (L) 7/1/13 – – Washington 11/3/98 11/2/08 (U) 11/3/98 (O) – 12/6/12 12/6/13 Wisconsin – – – – – 12/19/03

Notes:. a Only states with laws or provisions effective as of December 31, 2014 are shown. Some states have amended laws or provisions since 2014; these changes are not

indicated in the table. b The effective date coded here reflects the commencement of legal protections for a patient to use marijuana medically. c U = universal home cultivation allowance, meaning that all registered patients can grow their own marijuana; P = permitted home cultivation, meaning that

registered patients also require a permit to grow their own marijuana. d L = dispensaries are legally licensed/registered and open for business; O = dispensaries are operating illegally or quasi-legally without state approval. e For the eleven states that passed marijuana decriminalization laws in the 1970 s prior to our study period, only the year of enactment is reported.

(footnote continued) misdemeanor criminal offense with no jail time to an infraction with a maximum $100 fine and no criminal record. Based on the definition of marijuana decriminalization (i.e., no possibility of incarceration upon conviction), the original 1976 legislation was used to code this law.

5 Most states have a zero tolerance policy for THC blood levels of 0 ng/ml but six states have nonzero thresholds of 1 ng/ml (PA), 2 ng/ml (NV, OH), and 5 ng/ml (CO, MT, WA). These differences are not accounted for in this study. Note that this operationalization does not include laws targeting specific groups, such as commercial truck drivers or drivers less than 21 years old.

6 To account for 2008 and 2014 unit missing data for West Virginia, we relied on a supplemental source (Reaves, 2011) and inverse distance weighted extrapolation, re- spectively. To adjust for the more widespread problem of partial reporting by law en- forcement agencies for a given state-year, officer counts were adjusted according the equation: (adjusted number of officers) = (reported number of officers) / [(UCR popu- lation coverage) / (Census population)], where (reported number of officers) is the UCR officer count based on reporting agencies, (UCR population coverage) is the estimated state population served by those reporting agencies, and (Census population) is the of- ficial state Census population estimate.

7 Annual reports are available for 1998-2014, which covers the majority of state-years with an active DRE program. Still, about 17% (190 of 1,122) of the state-year observa- tions were missing on this measure, so we linearly interpolated these data between each state’s DRE program inception date and 2014, when all 50 states and DC reported into the program.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

59

measures were derived for narcotic+, depressant+, stimulant+, and other illicit+ (i.e., hallucinogens, PCP, anabolic steroids, inhalants) drugs. BAC (blood alcohol content) was operationalized as a three-level ordinal variable reflecting cutoffs for any alcohol use and the current legal driving limit: (1) BAC = 0, (2) 0 < BAC < 0.08, (3) BAC ≥ 0.08. Additional driver characteristics include race (white = 1, non- white = 0), gender (male = 1), age, and, given evidence of an inverse association between driver weight and BAC (Dunn and Tefft, 2014), driver BMI (body mass index). Other driver-level risk factors include the following: prior driving incident within the previous three years (i.e., vehicle crash, license suspension or revocation, DWI conviction, speeding conviction, or other harmful moving violation or conviction); speeding driver as evidenced by a charged offense or information in the police accident report, distracted driver due to diversion or inattention (e.g., texting, eating); current invalid license; safety device misuse defined as the misuse or nonuse of a seatbelt or helmet; and failure to obey road rules such as following improperly, failing to keep in proper lane, or passing where prohibited.

Relevant contextual information was also coded for each crash event. Adverse driving conditions reflects either the presence of bad weather (e.g., rain, sleet, snow, fog) or an unfavorable road surface (e.g., ice, sand, oil). To account for high-risk driving times, weekend accident indicates whether the crash occurred between 6 p.m. Friday and 5:59 a.m. Monday, nighttime accident indicates whether the crash occurred between 6 p.m. and 5:59 a.m. on any day of the week, and holiday accident codes whether the accident occurred during set na- tional holidays (i.e., New Year’s, Memorial Day, Fourth of July, Labor Day, Thanksgiving, and Christmas). Urban setting identifies whether the accident occurred in an urban versus rural setting. Additionally, single vehicle crash denotes whether the crash involved only the driver’s ve- hicle, and fatally injured driver indicates whether the driver died from injuries sustained in the accident.

Several additional variables control for potential background con- founders. State-level crash risk exposure is measured by per capita VMT (vehicle miles travelled, in thousands) (Federal Highway Administration, 2016). Although imperfect, the final four controls ac- count for regional differences (i.e., Northeast, Midwest, South, West) in background substance use and marijuana market characteristics, as state-level measures were either unavailable or unreliable. Alcohol prevalence (30-day) and marijuana prevalence (annual) are taken from the Monitoring the Future (MTF) survey of 12th grade students (Johnston et al., 2014). As a proxy for broader population-level trends in substance use, MTF has the advantage of providing consistent and unbroken data series for these measures. Lastly, mean THC% and sin- semilla % are regional estimates of average marijuana potency and the market share captured by high-quality seedless marijuana, respectively, according to federal seizures submitted to NIDA’s Potency Monitoring Program (e.g., ElSohly et al., 2016).

2.2. Empirical approach

This study employs a differences-in-differences (DiD) design, which estimates policy effects by comparing changes in the target outcome pre- and post-law in policy adopting states to respective changes in states that did not enact the policy (Angrist and Pischke, 2009; Bertrand et al., 2004; Meyer, 1995; Wooldridge, 2010). Due to the binary out- come, estimation was performed using the generalized linear model (GLM) with a binomial distribution and logit link function (Athey and Imbens, 2006; Lechner, 2011; Puhani, 2012; Strumpf et al., 2017). Multiple imputation was employed to handle missing data in a statis- tically robust manner (Little and Rubin, 2002). Various sensitivity analyses of this specification are performed. All analyses were con- ducted using Stata MP 15.0.

2.2.1. Multiple imputation of missing data The amount of missing data on alcohol and drug testing outcomes in

FARS is substantial, amounting, respectively, to 53.0% and 72.5% for the 1993–2014 study period. Multiple imputation (MI) is a state-of-the- art methodology for addressing missing data to obtain unbiased in- ferences. MI works by fitting a model to observed data to estimate missing data, generating multiple imputed datasets, m, that are each analyzed to produce results, which are then combined according to statistical rules to produce point estimates and standard errors that account for estimation uncertainty due to missingness (Rubin, 1987). Since 2001, NHTSA has multiply imputed missing BAC values in FARS and distributed the data files for analysis by the research community (Rubin et al., 1998; Subramanian, 2002). McGinty et al. (2017), for instance, recently used these imputed datasets to investigate the effects of ignition interlock laws on alcohol-involved fatal motor vehicle cra- shes.

To achieve valid and unbiased inference within an MI framework, a standard assumption is that the data are missing at random (MAR)—that is, missingness depends on the observed data but is in- dependent of any unobserved data. Although some commentators have questioned whether FARS drug testing data can be considered MAR (Slater et al., 2016), there is no direct test of this assumption. For- tunately, the incorporation into the imputation model of auxiliary variables that correlate with missingness on the outcome can help make the MAR assumption more plausible (Collins et al., 2001; Sullivan et al., 2015). This strategy was adopted here by including two sets of auxiliary variables in the MI prediction model. First, I use police reported alcohol involvement and police reported drug involvement, which record the re- sponding officer’s behavioral assessment of driver substance use, under the assumption that toxicological testing is more likely when the re- sponding officer suspects driver impairment. Second, the measures blood test for alcohol and blood test for drugs, which indicate whether the driver received a blood test for intoxicating substances, are used as auxiliary measures because blood testing is a recommended procedure over urinalysis in driving under the influence cases due to the method’s sensitivity to recent versus long-term use (Farrell et al., 2007; Logan et al., 2013).8 Overall, the auxiliary variables are fully observed for 24.0% of the cases, with 27.0% missing a single item, 34.7% missing two, 10.3% missing three, and 4.0% missing all four. In other words, 51.0% of cases have missing data on no more than one auxiliary vari- able and 85.7% contain missingness on no more than two variables. Although there is little guidance in the literature on the use of auxiliary variables that suffer from missing data themselves, I perform a sensi- tivity analysis on this below.

To assess each auxiliary variables’ correlation with the pattern of missingness on the outcome (missing/no drug test = 0, valid positive/ negative drug test = 1), I estimated a series of preliminary GLM models with a binomial distribution and logit link function, incorporating both state and year fixed effects with SEs clustered by state. These analyses confirm that each auxiliary variable is significantly associated with the probability of observing the outcome: police reported drug involvement (P = 0.40, N = 379,977, p < 0.001), police reported alcohol involve- ment (P = 0.19, N = 777,265, p < 0.001), blood test for drugs (P = 0.84, N = 1,020,767, p < 0.001), and blood test for alcohol (P = 0.57, N = 833,911, p < 0.001). Thus, including these auxiliary measures in the multiple imputation model will, on balance, increase the plausibility of the MAR assumption.

Notably, an added benefit of using auxiliary variables is that they can improve the quality of predictions to the extent they are correlated with the response on the incomplete outcome (Sullivan et al., 2015). A series of similar models predicting cannabis positivity among cases with observed data shows that each auxiliary variable is significantly asso- ciated with the probability of the driver testing positive for cannabis: police reported drug involvement (P = 0.36, N = 125,692, p < 0.001),

8 Oral fluid testing is also recommended for determining proximal versus distal drug use, but this matrix is not common nor is it coded independently in the FARS data.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

60

police reported alcohol involvement (P = 0.08, N = 201,632, p < 0.001), blood test for drugs (P = -0.05, N = 315,723, p < 0.05), and blood test for alcohol (P = -0.04, N = 265,920, p < 0.001).

Overall, these preliminary analyses suggest that including these four auxiliary variables in the MI prediction model will not only increase the plausibility that the data are MAR, but also reduce imputation bias and inefficiency (Collins et al., 2001; Sullivan et al., 2015). Importantly, for this research, such gains appear to hold even in the face of extreme levels of missingness on the outcome (Kontopantelis et al., 2017).

With this justification, multiple imputation by chained equations (MICE) was employed to fill in missing values for variables in the da- taset via a sequence of conditional specifications (Royston and White, 2011; White et al., 2011). Specifically, MICE employs an iterative procedure in which, initially, all missing values for x1, …, xk are ran- domly sampled from their respective distributions of observed values. Missing values for the first variable, x1, are then drawn from the pos- terior predictive distribution corresponding to a regression of x1 on x2, …, xk for observed cases of x1. This process is repeated for all variables with missing values using a burn-in cycle of 10 to improve stability of results before a single imputed data set, m, is produced. For this study, m = 100 datasets were imputed using a bootstrap approach for draws from the posterior predictive distribution. Per standard practice, all substantive and auxiliary variables, including the dependent variable and state and year fixed effects, were included in the MICE models, with regression specifications based on either linear regression (for con- tinuous variables) or binary or ordinal logistic regression (for catego- rical and ordinal variables) (StataCorp, 2015).

The standard MICE model was customized in two respects to ad- dress certain idiosyncrasies in the data. First, because FARS collects data on race only for fatally injured drivers, injury severity was selec- tively omitted as a predictor when imputing race, and vice versa, to avoid the problem of perfect prediction and facilitate model con- vergence. Second, due to the extreme variation in state drug testing rates (Logan et al., 2013), state fixed effects were swapped for regional fixed effects when imputing the drug testing outcomes in order to cluster these predictions across states with similar economic, political, and cultural attributes as defined by the eight U.S. Bureau of Economic Analysis regions.9

Turning to estimation, the GLM estimates were combined from analyses of the m = 100 imputed datasets applying Rubin’s combina- tion rules (Rubin, 1987; StataCorp, 2015).10 The standard MI approach of retaining cases with an imputed outcome during estimation was employed rather than the alternative method of “multiple imputation then deletion” (MID), which drops these cases prior to estimation (Von Hippel, 2007), because simulation research indicates that MID produces biased results when the outcome is correlated with auxiliary variables, and this bias only worsens as both the amount of missing data and the correlation between the auxiliary and outcome variables increases (Johnson and Young, 2011; Kontopantelis et al., 2017; Sullivan et al., 2015). Various sensitivity analyses of this preferred missing data spe- cification are performed below.

2.2.2. Differences-in-differences model The differences-in-differences specification follows:

+ = ∝ + + + + + + +βX τ ηP ωZ γ θ εCannabis MMList ist st st st s t ist (1)

where +Cannabis ist is the indicator for cannabis positivity and Xist is the vector of driver-level and contextual factors for driver i in state s at time t. The focal policy indicators, MMLst, capture whether a medical mar- ijuana law and its specific supply provisions were operational in state s at time t. Rival policy variables for state s at time t are captured in Pst. The vector Zst includes state/regional controls in jurisdiction s at time t. Finally, state fixed effects (γs) and year fixed effects (θt) control for potential unobserved confounders that are invariant across jurisdiction and time. In all models, SEs are clustered by state to ensure proper inference (Bertrand et al., 2004; Primo et al., 2007). Given the binary outcome, model estimation is performed by GLM with a logit index structure, ln[μ/(1-μ)], with the dependent variable distributed as Ber- noulli (Hardin and Hilbe 2012).

3. Results

Table 2 presents imputed descriptive statistics for the population of drivers involved in fatal vehicle accidents during the period 1993–2014, including the percentage of cases with imputed missing values. More fatal crash drivers tested positive for cannabis (8.8%) than other drugs, including stimulants (6.1%), narcotics (4.2%), depressants (3.6%), and other illicit drugs (0.3%). About one-quarter of fatal crash drivers tested positive for alcohol, greater than four-fifths of whom were above the BAC limit of .08. Most drivers tend to be male (74%) and white (74%) and are, on average, 40 years old and somewhat overweight (BMI = 26). Among drivers involved in fatal accidents, many had at least one prior driving incident (44%), failed to obey rules of the road (39%), or misused a safety belt or helmet (36%). About one- fifth were speeding (21%), with relatively fewer drivers operating with an invalid license (12%) or driving distracted (9%). Roughly four in ten of these drivers were fatally injured (46%) or involved in a single ve- hicle crash (38%). Other contextual factors indicate nearly half the accidents occurred at night (46%), in an urban setting (44%), or over the weekend (40%). Comparatively fewer accidents occurred in adverse conditions (18%) or during holiday periods (6%).

Concerning state-level marijuana policies, 28% of drivers involved in a fatal crash operated in a decriminalized state, followed by states with THC per se laws (16%), medical marijuana laws (14%), and re- creational marijuana laws (0.2%). About 12% of drivers operated in MML states with home cultivation provisions, and 11% operated in MML states with dispensaries (evenly split between states with licensed versus unlicensed dispensaries). For other policies, 69% of drivers op- erated in states with a BAC limit of 0.08, there was a mean of 237 sworn officers per 100,000, and there were about 2 DREs per 100,000 on average. The auxiliary variables indicate that 24% of responding offi- cers reported evidence of driver alcohol involvement, and 12% sus- pected drug involvement. With respect to toxicological screening, one- third to one-half of the drivers involved in a fatal accident from 1993 to 2014 received a blood test, respectively, for drugs or alcohol.

Table 3 presents the differences-in-differences estimates of the ef- fects of MMLs on cannabis-positive driving. Results indicate that med- ical marijuana laws in general are not significantly associated with marijuana-involved driving among the population of drivers involved in a fatal vehicle accident. Moreover, with respect to specific medical marijuana supply provisions, neither home cultivation nor unlicensed dispensaries are significantly linked to the likelihood of marijuana-in- volved driving. Only in states with licensed medical marijuana dis- pensaries is there a significant increase in the odds of marijuana-in- volved driving of 14%.

To apply a more concrete causal interpretation to these results, I estimate the average treatment effect on the treated (ATET) for two relevant policy contrasts. The ATET, which corresponds to the effect among states that actually passed an MML, is the quantity of interest because MMLs are not universal treatments randomly adopted by the states. The first policy contrast indicates that drivers in MML states with licensed dispensaries have, on average, a 0.014 greater probability

9 These regions include New England (CT, ME, MA, NH, RI, VT), Mideast (DE, DC, MD, NJ, NY, PA), Great Lakes (IL, IN, MI, OH, WI), Plains (IA, KS, MN, MO, NE, ND, SD), Southeast (AL, AR, FL, GA, KY, LA, MS, NC, SC, TN, VA, WV), Southwest (AZ, NM, OK, TX), Rocky Mountains (CO, ID, MT, UT, WY), and Far West (AK, CA, HI, NV, OR, WA).

10 Two measures were transformed between the MICE prediction and estimation phases of the analysis. First, home cultivation was collapsed from a trichotomous to a dichotomous variable because only two states required home growing permits during the study period. Second, to improve prediction of missing data, the highly skewed measure DREs per 100,000 was initially logged but then retransformed to levels for estimation.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

61

(p < 0.01) of testing positive for cannabis compared to the counter- factual condition of no state MML. The second policy contrast, which compares drivers in MML states with licensed dispensaries to the counterfactual condition of MML states without licensed or unlicensed dispensaries, indicates a .011 greater probability (p < 0.05) of testing positive for cannabis. Although significant, these are small effects. A back-of-the-envelope calculation suggests that of the roughly 8000 fatal-crash-involved drivers from licensed dispensary states in 2014, an additional 87 to 113 tested positive for marijuana than otherwise would have due to the dispensary policy (8057 * .011 ≈ 87; 8057 * 0.014 ≈ 113).

Table 3 also reveals that several rival policy variables impact can- nabis-involved driving. Marijuana decriminalization is associated with

a 24% increase in the odds of cannabis-involved driving, although this effect is only significant at the p < 0.10 level. However, the lack of policy variation for eleven states that passed decriminalization in the 1970 s prior to the study period likely reduces estimation precision. Conversely, recreational marijuana laws are associated with an 18% decrease in the odds of cannabis-involved driving. This effect might be explained by heightened post-legalization drugged driving enforce- ment, a proposition that finds some support in the literature (Urfer et al. 2014), although this effect should be treated with caution as it is identified from for a one-year post-law period in just two states (CO and WA). Notably, both the THC and BAC .08 per se laws significantly re- duce the odds of cannabis-positive driving by 0.18 and 0.21, respec- tively. The latter effect suggests that drunk driving laws have com- plementary deterrent effects on cannabis-involved driving. Lastly, the number of police officers per 100,000—but not the number of DREs per 100,000—is associated with lower odds of cannabis-positive driving. In particular, a standard deviation increase in police officers per 100,000

Table 2 Imputed descriptive statistics, 1993–2014.

Variables Percent/mean (SD) Percent imputed

Dependent Variable Cannabis+ 8.8 72.5 Focal Medical Marijuana Laws Medical Marijuana 14.3 0.0 Home Cultivation 12.3 0.0 Dispensaries 0.0 None 89.4 Unlicensed 5.4 Licensed 5.2

Rival State Policies Marijuana Decriminalization 28.0 0.0 Recreational Marijuana 0.2 0.0 THC Per Se Law 15.9 0.0 BAC .08 Law 68.5 0.0 Police Officers per 100,000 236.9 (50.0) 0.0 DREs per 100,000 1.6 (1.9) 0.0 Driver-Level Factors Narcotic+ 4.2 72.5 Depressant+ 3.6 72.5 Stimulant+ 6.1 72.5 Other Illicit+ 0.3 72.5 BAC 53.0 BAC=0 75.1 0 < BAC < 0.08 4.7 BAC ≥ 0.08 20.1

White 73.6 71.9 Male 73.9 1.4 Age 40.2 (18.0) 1.6 BMI 26.2 (5.3) 45.9 Prior Driving Incident 44.2 6.2 Speeding Driver 21.0 2.2 Distracted Driver 8.9 4.0 Invalid License 11.6 2.4 Safety Device Misuse 36.1 9.9 Failure to Obey Road Rules 38.9 2.1 Contextual Factors Adverse Driving Conditions 17.9 0.6 Weekend Accident 39.5 0.6 Nighttime Accident 46.1 0.6 Holiday Accident 6.4 0.6 Urban Setting 44.3 0.5 Single Vehicle Crash 38.4 0.0 Fatally Injured Driver 46.0 1.3 Background Factors Per Capita VMT (Thousands) 10.0 (1.5) 0.0 Mean THC % 7.5 (3.3) 0.0 Sinsemilla % 23.4 (26.0) 0.0 Alcohol Prevalence (30-Day) 46.1 (5.6) 0.0 Marijuana Prevalence (Annual) 33.9 (4.2) 0.0 Auxiliary Variables Police Reported Alcohol Involvement 23.9 33.7 Police Reported Drug Involvement 11.8 67.6 Blood Test for Alcohol 50.1 28.9 Blood Test for Drugs 32.8 13.0

Note: N = 1,171,261 reflects the average analytic sample across the m = 100 imputed datasets, which varies between 1,171,175 and 1,171,339 because age was imputed for 1.6% of cases.

Table 3 Differences-in-differences estimates of the effect of medical marijuana laws on marijuana-involved driving among drivers involved in fatal crashes, 1993–2014.

Variables Coefficient SE OR [95% CI]

Focal Medical Marijuana Laws Medical Marijuana 0.052 0.064 1.05 [0.93, 1.19] Home Cultivation 0.175 0.185 1.19 [0.83, 1.71] Dispensaries Unlicensed 0.065 0.084 1.07 [0.91, 1.23] Licensed 0.134 0.061* 1.14 [1.02, 1.29]

Rival State Policies Marijuana Decriminalization 0.212 0.125† 1.24 [0.97, 1.58] Recreational Marijuana −0.197 0.078* 0.82 [0.70, 0.96] THC Per Se Law −0.123 0.047** 0.88 [0.81, 0.97] BAC 0.08 Law −0.235 0.047*** 0.79 [0.72, 0.87] Police Officers per 100,000 −0.002 0.001* 1.00 [1.00, 1.00] DREs per 100,000 −0.008 0.010 0.99 [0.97, 1.01] Driver-Level Factors Narcotic+ −0.033 0.034 0.97 [0.90, 1.04] Depressant+ 0.353 0.035*** 1.42 [1.33, 1.52] Stimulant+ 0.541 0.032*** 1.72 [1.61, 1.83] Other Illicit+ 0.480 0.077*** 1.62 [1.39, 1.88] BAC 0 < BAC < 0.08 0.483 0.025*** 1.62 [1.54, 1.70] BAC ≥ 0.08 0.303 0.026*** 1.35 [1.29, 1.42] White 0.198 0.022*** 1.22 [1.17, 1.27] Male 0.457 0.019*** 1.58 [1.52, 1.64] Age −0.038 0.001*** 0.96 [0.96, 0.96] BMI −0.024 0.002*** 0.98 [0.97, 0.98] Prior Driving Incident 0.205 0.013*** 1.23 [1.20, 1.26] Speeding Driver 0.108 0.018*** 1.11 [1.08, 1.15] Distracted Driver −0.027 0.025 0.97 [0.93, 1.02] Invalid License 0.246 0.020*** 1.28 [1.23, 1.33] Safety Device Misuse 0.243 0.017*** 1.27 [1.23, 1.32] Failure to Obey Road Rules −0.063 0.019*** 0.94 [0.90, 0.97] Contextual Factors Adverse Driving Conditions −0.041 0.018* 0.96 [0.93, 0.99] Weekend Accident −0.039 0.014** 0.96 [0.93, 0.99] Nighttime Accident −0.024 0.016 0.98 [0.94, 1.01] Holiday Accident 0.014 0.025 1.01 [0.97, 1.07] Urban Setting −0.037 0.019† 0.96 [0.93, 1.00] Single Vehicle Crash 0.010 0.016 1.01 [0.98, 1.04] Fatally Injured Driver −0.336 0.044*** 0.71 [0.66, 0.78] Background Factors Per Capita VMT (Thousands) 0.032 0.026 1.03 [0.98, 1.09] Mean THC % 0.047 0.009*** 1.05 [1.03, 1.07] Sinsemilla % 0.012 0.002*** 1.01 [1.01, 1.02] Alcohol Prevalence (30-Day) 0.010 0.006† 1.01 [1.00, 1.02] Marijuana Prevalence (Annual) 0.011 0.006† 1.01 [1.00, 1.02]

Note: Estimates obtained by GLM with a logit link function and binomial dis- tribution within MICE framework for m = 100 datasets. Model includes state and year fixed effects, with standard errors clustered by state. N = 1,171,175; † p < .10; * p < .05; ** p < .01; *** p < .001.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

62

decreases these odds by a factor of 0.90 (e−.002*50.0). For the interested reader, Table 3 also reports results for driver-level, contextual, and background factors.

3.1. Model robustness and identification checks

Table 4 presents a series of robustness checks of the multiple im- putation procedure across a series of alternate specifications/sub- samples. Model 1 restricts the analysis to cases in which at least two auxiliary variables were fully observed. Model 2 implements the MICE approach using ‘multiple imputation then deletion’ (Von Hippel, 2007), which drops cases from the estimation model for which the outcome was imputed. Model 3 is estimated on the subpopulation of fatally in- jured drivers, who are systematically tested at a much higher rate than surviving drivers. Model 4 includes only post-2001 years for which drug testing rates are also systematically higher. Model 5 includes only states with cumulative drug testing rates above the overall state median of 33.4%. Finally, Model 6 focuses only on drivers who received a blood test for drugs, which is more likely to capture recent drug use compared to other screening methods. Overall, these six alternate specifications produce policy effects that are consistent with the base model reported in Table 3. The licensed dispensary effect remains significant across all models (although two models are significant only at the p < 0.10 level). Moreover, with a single exception for the main medical marijuana law indicator in Model 4, no other policy effects reach significance, which is again consistent with the base model. In short, these robustness checks increase confidence in the underlying imputation model.

Table 5 reports various identification checks of the DiD model. In panel A, Model 1 again reproduces the base specification from Table 3 for comparison. Model 2 adds state-specific linear trends, and Model 3 includes only supply provisions to remove potential multicollinearity with the main MML policy indicator. In both cases, model coefficients and standard errors are similar to the base model, indicating the find- ings are robust to these alternate specifications. Panels B and C examine two key assumptions of DiD estimation: (i) the existence of common

trends in the outcome between the treatment and comparison groups in the pre-intervention period and (ii) the absence of post-intervention spillover from the treatment group to the comparison group (Ryan et al., 2015). The common trends assumption was assessed by adding three lead policy indicators to the base model for the set of MML policy indicators. A pattern of results in which the coefficients on the leads are not significantly different from zero is consistent with the common trends assumption, and this pattern is confirmed by the results reported in panel B. Spillover effects are assessed in panel C of Table 5 by in- cluding an analogous set of policy indicators for non-MML border states (i.e., neighboring state law). The results indicate that cannabis-involved driving among drivers involved in fatal vehicle accidents is significantly less likely in neighboring states, which may reflect a deterrent effect due to an increased law enforcement response in border communities (see e.g., Ellison and Spohn, 2015). Thus, if the observed legal dispensary effect on cannabis-positive driving is partly identified from increased drugged-driving deterrence in neighboring states, the reported dis- pensary effect is potentially biased and might reflect a false-positive policy finding. Future research should more directly investigate spil- lover effects in this policy space and assess the attendant implications for causal inference (e.g., Grogger, 2002).

Table 4 Robustness checks across alternate specifications and subsamples.

Policy indicators

(1) Two or more auxiliary measures observed

(2) Multiple imputation then deletion

(3) Fatally injured drivers

Medical Marijuana

0.059 (0.065) 0.104 (0.108) 0.016 (0.122)

Home Cultivation

0.179 (0.183) 0.123 (0.187) 0.163 (0.176)

Dispensaries Unlicensed 0.037 (0.078) −0.009 (0.156) 0.067 (0.166) Licensed 0.124 (0.056)* 0.417 (0.107)*** 0.365 (0.125)**

N = 1,004,453 States = 51, Years = 22

N = 322,458 States = 51, Years = 22

N = 538,115 States = 51, Years = 22

Policy indicators (4) Post-2001 years

(5) High-rate testing states

(6) Blood drug testing

Medical Marijuana

0.130 (0.053)* 0.064 (0.094) 0.014 (0.122)

Home Cultivation 0.204 (0.170) 0.180 (0.191) 0.238 (0.164) Dispensaries Unlicensed −0.021 (0.057) 0.018 (0.096) 0.004 (0.119) Licensed 0.104 (0.062)† 0.131 (0.073)† 0.385 (0.106)***

N = 666,279 States = 51, Years = 13

N = 555,083 States = 26, Years = 22

N = 383,730 States = 51, years = 22

Note: Reported statistics are the log odds and standard errors clustered by state. Each model includes state and year fixed effects and all control variables. Results reflect MICE estimation on m = 100 datasets, delimited by the sub- samples indicated. † p < .10; * p < .05;** p < .01; *** p < .001.

Table 5 Assessing identification of the differences-in-differences model.

Panel A: Alternate Specifications

Policy indicators (1) State trends (2) Supply provisions only Medical Marijuana 0.080 (0.058) – Home Cultivation 0.219 (0.180) 0.178 (0.184) Dispensaries Unlicensed 0.040 (0.075) 0.102 (0.083) Licensed 0.195 (0.094)* 0.168 (0.060)**

Panel B: Common Trends

Policy indicators Year of law change Two-year lead Medical Marijuana 0.047 (0.080) 0.038 (0.587) Home Cultivation 0.180 (0.183) −0.066 (0.068) Dispensaries Unlicensed 0.108 (0.082) −0.021 (0.070) Licensed 0.187 (0.080)* −0.030 (0.069)

One-year lead Three-year lead Medical Marijuana 0.078 (0.094) −0.025 (0.051) Home Cultivation −0.115 (0.123) 0.106 (0.068) Dispensaries Unlicensed −0.031 (0.059) 0.007 (0.055) Licensed −0.035 (0.073) 0.000 (0.070)

Panel C: Spillover Effects

Policy indicators Own state law Neighboring state law Medical Marijuana 0.069 (0.069) −0.060 (0.091) Home Cultivation 0.176 (0.182) 0.103 (0.100) Dispensaries Unlicensed 0.016 (0.080) −0.105 (0.074) Licensed 0.077 (0.064) −0.109 (0.048)*

Note: Reported statistics are the log odds and standard errors clustered by state. Each model includes state and year fixed effects and all control variables. Results reflect MICE estimation on m = 100 datasets. Panel A presents alter- native specifications that add controls for individual state trends (Model 1) and drop the generic MML indicator to gauge potential multicollinearity (Model 2). Panel B provides a test of the common trend assumption by estimating a single model of the contemporaneous policy effect together with three annual lags. A lack of significance on the lagged policy variables is consistent with the common trend assumption. Panel C provides a test of spillover effects by esti- mating a single model with policy indicators for states with their own medical marijuana law provisions together with policy indicators denoting states ad- jacent to states with the respective medical marijuana law provisions. A sig- nificant neighboring state law effect is consistent with a spillover effect. † p < .10; * p < .05; ** p < .01; *** p < .001.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

63

4. Discussion

The proliferation of medical marijuana laws in the US has raised concerns about the potential adverse effects on roadway safety (Turnbull & Hodge, 2017). The fear is that expanded availability of more potent strains of marijuana coupled with lax norms on drugged driving will increase the prevalence of driving while high (Davis et al., 2016), and that this will result in an increased number of fatal and serious crashes. Using driver-level FARS data from all 51 states over the years 1993–2014, this study implemented a differences-in-differences analysis within a multiple imputation framework to investigate the causal effect of state medical marijuana laws on cannabis-positive driving among the universe of drivers involved in a fatal vehicle acci- dent. In addition to examining the general impact of state medical marijuana laws, the study also sought to understand the effects of specific marijuana supply provisions (i.e., home cultivation, dis- pensaries) on cannabis-involved driving. Specific methodological in- novations of this study include the use of driver-level data, application of multiple imputation to address missing data, and policy coding that captures variations in state medical marijuana laws.

Despite reporting positive coefficients on all MML policy indicators, the results of this study indicate that medical marijuana laws in general have null effects on the prevalence of cannabis-positive driving. The key exception is for MML states that regulate the sale of cannabis though dispensaries, a policy framework that was shown to increase the probability of cannabis-positive driving by .011–.014, depending on the counterfactual policy. However, as noted above, this is a relatively small effect, representing an additional 87–113 cannabis-positive dri- vers in 2014 who were involved in fatal vehicle accidents who might not otherwise have been.

How, then, might dispensaries increase the incidence of cannabis- positive driving? Existing incentives (e.g., price, selection) for medical marijuana patients to travel to these supply centers appear to play a key role, and such incentives only increase with legal or practical con- straints on other sources of supply (e.g., home cultivation prohibited, black market uncertainty). The expansion of medical marijuana pro- grams, along with point-of-sale quantity limitations and geographical restrictions on dispensary locations, would further increase the fre- quency and distance of such travel. States also vary in the qualifying conditions permitting access to medical marijuana programs. To the extent that a state’s patient population is characterized less by infirm individuals with serious medical conditions, who tend to drive less, the conditional risk of cannabis-involved driving will be higher.

Although state dispensary regulations uniformly prohibit the onsite consumption of marijuana, if dispensary customers subsequently use marijuana proximal to driving, then dispensaries can plausibly be linked to the incidence of cannabis-positive driving. If this translates into a greater risk of cannabis-impaired driving, then we should observe a concomitant increase in the probability of cannabis-involved fatal traffic accidents. Although small in magnitude, this study documents just such a policy effect. It is noteworthy that an analogous policy effect was not detected for MML states with unlicensed dispensaries. This is not surprising given the uncertainty and volatility of such policy en- vironments, where dispensary operators face an uncertain regulatory environment and greater enforcement action.

Despite employing a sophisticated missing data approach and a rigorous quasi-experimental differences-in-differences design, study limitations may moderate a causal interpretation of these findings. First, testing positive for cannabis does not necessarily imply impair- ment (Berning and Smither, 2014). Inactive metabolites of THC, the main psychoactive component of marijuana, can be detected in urine samples for several weeks or longer following marijuana use. To the extent that FARS captures distal rather than proximal marijuana use, the causal link between medical marijuana policies, impaired driving, and crash risk becomes less robust. Importantly, a sensitivity analysis conducted on the subsample of drivers who received a blood test for

drugs, which has a cannabinoid detection window of hours and days rather than weeks and months, reveals a consistent set of results that allays this concern. Second, absent national standards, substantial variation exists across both states and time in drug testing protocols for suspected driver impairment (Berning and Smither, 2014). Future re- search, both retrospective and prospective in design, is needed to better understand the policy-relevant implications of state-level variations in roadside drug testing procedures. Third, FARS suffers from an in- credible amount of missing data on drug testing outcomes. Analyzing only the observed data would likely produce biased and inefficient re- sults. Although some observers have expressed skepticism that FARS drug testing data can be fruitfully used in policy research (Romano et al., 2017; Slater et al., 2016), this study implemented a rigorous multiple imputation procedure that was robust to several alternate specifications, thereby increasing confidence in the core findings. However, if there are systematic tendencies that are not captured by the auxiliary variables and other covariates, the imputation and results may be biased. Fourth, a sensitivity analysis assessing spillover effects, which informs identification of the DiD model, showed that cannabis- positivity among drivers involved in fatal accidents was significantly lower in states that border MML states with licensed dispensaries. This reverse spillover effect is consistent with increased drugged driving enforcement in bordering states (Ellison & Spohn, 2015), which lessens confidence that the observed effects can be wholly attributed to medical marijuana dispensary policies.

Even if limited to a conditional association interpretation, this study’s findings have several policy implications. First, for states con- sidering adopting a dispensary model, consideration should also be given to prophylactic roll out of evidence-based prevention programs that target lenient attitudes on marijuana use and driving. Shaping drug use and driving norms and perceptions of risk prior to a dispensary program’s unveiling is likely to be more effective than any post hoc targeting of user attitudes and behaviors. Second, dispensary access is uneven in many states because of local prohibitions, so regulatory agencies that rule on dispensary applications and locations might weigh projected demand and travel time from underserved populations when making dispensary siting decisions. Third, states should allow and regulate marijuana delivery services in the interest of roadway safety (Freisthler and Gruenewald, 2014). Fourth, this study found that THC per se laws significantly reduced the likelihood of cannabis-positive driving. Thus, coupling principled THC per se laws (see, e.g., Huestis, 2015) with state dispensary laws may provide an effective counter to the potential negative externalities of dispensaries.

Future research can expand upon this study in several ways. First, studies that more directly examine potential spillover effects will in- crease our understanding of how these laws affect neighboring states and vice versa. Second, this study examined variations in medical marijuana supply policies, but there are also important state differences in the number and type of medical marijuana patients. Investigating the effects of MMLs on drugged driving while controlling for program size and/or patient characteristics would provide additional policy insights. Finally, performing spatial analyses that examine the geographic asso- ciation between dispensary locations and marijuana-related driving infractions would be fruitful. Notably, FARS data provide geographic coordinates for fatal accidents, but investigators should also explore the use of other georeferenced data sources. State marijuana laws continue to evolve, so rigorous policy research in this area must keep pace.

References

Anderson, D..Mark, Hansen, Benjamin, Rees, Daniel I., 2013. Medical marijuana laws, traffic fatalities, and alcohol consumption. J. Law Econ. 56 (2), 333–369.

Angrist, Joshua D., Pischke, Jörn-Steffen, 2009. Mostly Harmless Econometrics: An Empiricist’S Companion. Princeton University Press, Princeton, NJ.

Athey, Susan, Imbens, Guido W., 2006. Identification and inference in nonlinear differ- ence‐in‐differences models. Econometrica 74 (2), 431–497.

Berning, Amy, Compton, Richard, Wochinger, Kathryn, 2015. Results of the 2013-2014

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

64

National Roadside Survey of Alcohol and Drug Use by Drivers. National Highway Traffic Safety Administration, Washington, DC.

Berning, Amy, Smither, Dereece D., 2014. Understanding the Limitations of Drug Test Information, Reporting, and Testing Practices in Fatal Crashes. National Highway Traffic Safety Administration, Washington, DC.

Bertrand, Marianne, Duflo, Esther, Mullainathan, Sendhil, 2004. How Much should we trust differences-in-differences estimates? The Q. J. Econ. 119 (1), 249–275.

Brady, Joanne E., Li, Guohua, 2014. Trends in alcohol and other drugs detected in fatally injured drivers in the United States, 1999-2010. Am. J. Epidemiol. 79 (6), 692–699.

Collins, Linda M., Schafer, Joseph L., Kam, Chi-Ming, 2001. A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychol. Methods 6 (4), 330–351.

Davis, Kevin C., Allen, Jane, Duke, Jennifer, Nonnemaker, James, Bradfield, Brian, Farrelly, Matthew C., Shafer, Paul, Novak, Scott, 2016. Correlates of marijuana drugged driving and openness to driving while High: evidence from Colorado and Washington. PLoS One 11 (1), 1–13.

Dunn, Richard A., Tefft, Nathan W., 2014. Has increased body weight made driving safer? Health Econ. 23 (11), 1374–1389.

Ellison, Jared M., Spohn, Ryan E., 2015. Borders Up in smoke: marijuana enforcement in Nebraska after Colorado’s legalization of medicinal marijuana. Crim. Just. Policy Rev. 28 (9), 847–865.

ElSohly, Mahmoud A., Mehmedic, Zlatko, Foster, Susan, Gon, Chandrani, Chandra, Suman, Church, James C., 2016. Changes in Cannabis potency over the last Two decades (1995-2014): analysis of current data in the United States. Biol. Psychiatry 79 (7), 613–619.

Farrell, Laurel J., Kerrigan, Sarah, Logan, Barry K., 2007. Recommendations for tox- icological investigation of drug impaired driving. J. Forensic Sci. 52 (5), 1214–1218.

Federal Bureau of Investigation, 1994-2015. Crime in the United States. Washington, DC: U.S. Department of Justice.

Federal Highway Administration, 2016. Highway Statistics Series 2016 [cited February 26 2016]. Available from http://www.fhwa.dot.gov/policyinformation/statistics. cfm.

Freisthler, Bridget, Gruenewald, Paul J., 2014. Examining the relationship between the physical availability of medical marijuana and marijuana use across fifty California cities. Drug Alcohol Depend. 143, 244–250.

Governor's Highway Safety Association. 2015. Drug Impaired Driving Laws 2015 [cited April 1 2015]. Available from http://www.ghsa.org/html/stateinfo/laws/dre_perse_ laws.html.

Grogger, Jeffrey, 2002. The effects of civil gang injunctions on reported violent crime: evidence from Los Angeles County. J. Law Econ. 45 (1), 69–90.

Hall, Wayne, 2015. What has research over the past Two decades revealed about the adverse health effects of recreational Cannabis use? Addiction 110 (1), 19–35.

Hamzeie, Raha, Thompson, Iftin, Roy, Sneha, Savolainen, Peter T., 2017. State-level comparison of traffic fatality data in consideration of marijuana laws. Transp. Res. Rec.: J.Transp. Res. Board (2660), 78–85.

Hardin, James W., Hilbe, Joseph, 2012. Generalized Linear Models and Extensions, 3rd ed. Stata Press, College Station, TX.

Huestis, Marilyn A., 2015. Deterring driving under the influence of Cannabis. Addiction 110 (11), 1697–1698.

International Association of Chiefs of Police. 1999-2015. The 2014 Annual Report of the Drug Recognition Expert Section. Alexandria, VA: Drug Recognition Expert (DRE) Section of the International Association of Chiefs of Police.

Johnson, David R., Young, Rebekah, 2011. Toward Best practices in analyzing datasets with missing data: comparisons and recommendations. J. Marriage Family 73 (5), 926–945.

Johnston, Lloyd D., O’Malley, Patrick M., Bachman, Jerald G., Schulenberg, J.E., Miech, R.A., 2014. Demographic Subgroup Trends Among Adolescents in the Use of Various Licit and Illicit Drugs, 1975-2013. The University of Michigan, Institute for Social Research, Ann Arbor, MI.

Klieger, Sarah B., Gutman, Abraham, Allen, Leslie, Liccardo Pacula, Rosalie, Ibrahim, Jennifer K., Burris, Scott, 2017. Mapping medical marijuana: State laws regulating patients, product safety, supply chains and dispensaries, 2017. Addiction 112, 2206–2216.

Kontopantelis, Evangelos, White, Ian R., Sperrin, Matthew, Buchan, Iain, 2017. Outcome- sensitive multiple imputation: a simulation study. BMC Med. Res. MEthodol. 17 (1), 1–13.

Lacey, John, Brainard, Katharine, Snitow, Samantha, 2010. Drug Per Se Laws: A Review of Their Use in States. National Highway Traffic Safety Administration, Washington, DC.

Lechner, Michael, 2011. The estimation of causal effects by difference-in-difference methods. Found Trends Econ. 4 (3), 165–224.

Little, Roderick J.A., Rubin, Donald B., 2002. Statistical Analysis With Missing Data. Wiley-Interscience, Hoboken, NJ.

Logan, Barry K., Kayla, J., Lowrie, Jennifer L., Turri, Jillian, K., Yeakel, Limoges, Jennifer F., Miles, Amy K., Scarneo, Colleen E., Kerrigan, Sarah, Farrell, Laurel J., 2013. Recommendations for toxicological investigation of drug-impaired driving and motor vehicle fatalities. J. Anal. Toxicol. 37, 552–558.

Marijuana Policy Project, 2016. State-by-State Medical Marijuana Laws: How to Remove the Threat of Arrest 2015, With a December 2016 Supplement. Marijuana Polcy Project, Washington, DC.

Masten, Scott V., Guenzburger, GloriamVanine, 2014. Changes in driver cannabinoid prevalence in 12 U.S. States after implementing medical marijuana laws. J. Saf. Res.

50, 35–52. McGinty, Emma E., Tung, Gregory, Shulman-Laniel, Juliana, Hardy, Rose, Rutkow,

Lainie, Frattaroli, Shannon, Vernick, Jon S., 2017. Ignition interlock laws: effects on fatal motor vehicle crashes, 1982-2013. Am. J. Prev. Med. 52 (4), 417–423.

Meyer, BreedD., 1995. Natural and quasi-experiments in economics. J. Bus. Econ. Stat. 13 (2), 151–161.

National Highway Traffic Safety Administration, 1998. Presidential Initiative for Making .08 BAC the National Legal Limit—Recommendations from the Secretary of Transportation. U.S. Department of Transportation, Washington, DC.

National Highway Traffic Safety Administration, 2015. Fatality Analysis Reporting System (FARS) Analytical User’S Manual 1975-2014. U.S. Department of Transportation, Washington, DC.

National Institute on Alcohol Abuse and Alcoholism, 2015. Blood Alcohol Concentration (BAC) Limits: Adult Operators of Noncommercial Motor Vehicles 2015 [cited April 1 2015]. Available from https://alcoholpolicy.niaaa.nih.gov/Blood_Alcohol_ Concentration_Limits_Adult_Operators_of_Noncommercial_Motor_Vehicles.html.

Pacula, RosalieLiccardo, Sevigny, Eric L., 2014. Marijuana liberalization policies: why we can’t learn Much from policy still in motion. J. Policy Anal. Manage. 33 (1), 212–221.

Primo, David M., Jacobsmeier, Matthew L., Milyo, Jeffrey, 2007. Estimating the impact of State policies and institutions with mixed-level data. State Polit. Policy Q. 7 (4), 446–459.

Puhani, Patrick A., 2012. The treatment effect, the Cross difference, and the interaction term in nonlinear “difference-in-differences” models. Econ. Lett. 115 (1), 85–87.

Reaves, Brian, 2011. Census of State and Local Law Enforcement Agencies, 2008. US Department of Justice, Office of Justice Programs., Washington, DC.

Rogeberg, Ole, Elvik, Rune, 2016. The effects of Cannabis intoxication on motor vehicle collision revisited and revised. Addiction 111, 1348–1359.

Romano, Eduardo, Pollini, Robin A., 2013. Patterns of drug use in fatal crashes. Addiction 108 (8), 1428–1438.

Romano, Eduardo, Torres-Saavedra, Pedro, Voas, Robert B., Lacey, John H., 2017. Marijuana and the risk of fatal car crashes: what can we learn from FARS and NRS data? J. Prim. Prevent. 38 (3), 315–328.

Royston, P., White, I.R., 2011. Multiple imputation by chained equations (MICE): im- plementation in stata. J. Statist. Softw. 45 (4), 1–20.

Rubin, Donald B., 1987. Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons, Inc, New York.

Rubin, Donald B., Schafer, Joseph L., Subramanian, Rajesh, 1998. Multiple Imputation of Missing Blood Alcohol Concentration (BAC) Values in FARS. National Highway Traffic Safety Administration, Washington, DC.

Rudisill, Toni M., Zhao, Songzhu, Abate, Marie A., Coben, Jeffrey H., Zhu, Motao, 2014. Trends in drug use among drivers killed in U.S. Traffic crashes, 1999-2010. Accid. Anal. Prev. 70, 178–187.

Ryan, Andrew M., Burgess, James F., Dimick, Justin B., 2015. Why we should not be indifferent to specification choices for difference‐in‐differences. Health Serv. Res. 50 (4), 1211–1235.

Salomonsen-Sautel, Stacy, Min, Sung-Joon, Sakai, Joseph T., Thurstone, Christian, Hopfer, Christian, 2014. Trends in fatal motor vehicle crashes before and after marijuana commercialization in Colorado. Drug Alcohol Depend. 140, 137–144.

Santaella-Tenorio, Julian, Mauro, Christine M., Wall, Melanie M., Kim, June H., Cerdá, Magdalena, Keyes, Katherine M., Hasin, Deborah S., Galea, Sandro, Martins, Silvia S., 2017. US traffic fatalities, 1985-2014, and their relationship to medical marijuana laws. Am. J. Public Health 107 (2), 336–342.

Scherer, Michael, Voas, Robert B., Furr-Holden, Debra, 2013. Marijuana as a predictor of concurrent substance use among motor vehicle operators. J. Psychoact. Drugs 45 (3), 211–217.

Slater, Megan E., Castle, I.J.enP., Logan, Barry K., Hingson, Ralph W., 2016. Differences in State drug testing and reporting by driver type in U.S. Fatal traffic crashes. Accid. Anal. Prev. 92, 122–129.

StataCorp, 2015. Stata Multiple-Imputation Reference Manual Release 14. Stata Press, College Station, TX.

Strumpf, Erin C., Harper, Sam, Kaufman, Jay S., 2017. Fixed effects and difference in differences. Methods in Social Epidemiology. Jossey-Bass, San Francisco, CA.

Subramanian, Rajesh, 2002. Transitioning to Multiple Imputation – a New Method to Impute Missing Blood Alcohol Concentration (BAC) Values in FARS. National Highway Traffic Safety Administration, Washington, DC.

Sullivan, Thomas R., Salter, Amy B., Ryan, Philip, Lee, Katherine J., 2015. Bias and precision of the “Multiple imputation, then deletion” method for dealing with missing outcome data. Am. J. Epidemiol. 182 (6), 528–534.

Turnbull, David, Hodge Jr, James G., 2017. Driving under the influence of marijuana laws and the public’s health: public health and the law. The J. Law, Med. & Ethics 45 (2), 280–283.

Urfer, Sarah, Morton, Jaime, Beall, Vanessa, Feldmann, Jeanna, Gunesch, Justin, 2014. Analysis of Δ9-tetrahydrocannabinol driving under the influence of drugs cases in Colorado from january 2011 to february 2014. J. Anal. Toxicol. 38 (8), 575–581.

Von Hippel, Paul T., 2007. Regression with missing ys: an improved strategy for analyzing multiply imputed data. Sociological Methodology 37 (1), 83–117.

White, I.R., Royston, P., Wood, A.M., 2011. Multiple imputation using chained equations: issues and guidance for practice. Stat. Med. 30 (4), 377–399.

Wilson, Fernando A., Stimpson, Jim P., Pagan, Jose A., 2014. Fatal crashes from drivers testing positive for drugs in the US, 1993-2010. Public Health Rep. 129 (4), 342–350.

Wooldridge, Jeffrey M., 2010. Econometric Analysis of Cross Section and Panel Data, 2nd ed. The MIT Press, Cambridge, MA.

E.L. Sevigny Accident Analysis and Prevention 118 (2018) 57–65

65