Medical Marijuana Laws and Suicide

Objective: To estimate the causal effect of a medical marijuana initiative on suicide risk. In 1996, California legalized marijuana use for medical purposes. Implementation was abrupt and uniform, presenting a “natural experiment.”<br><br>Method: Total, gun and non-gun suicides were aggregated by state for the years 1970-2004. California’s control time series was constructed as a weighted combination of the 41 states that did not legalize marijuana during the time-frame. Post-intervention differences for California and its constructed control time-series were interpreted as the effects of the medical marijuana law on suicide. Significance of the effects were assessed with permutation tests.<br><br>Results: The 1996 legalization resulted in mean annual reductions of 398.9 total suicides, 208 gun suicides, and 135 non-gun suicides. The effect estimates for total and gun suicides were statistically significant (p&lt;.05) but the effect estimate for non-gun suicides was not (p≥.488).<br><br>Conclusions: Since the effect for non-gun suicides was indistinguishable from chance, we infer that the overall causal effect was realized through gun suicides. The mechanism could not be determined, however. Participation in the medical marijuana program legally disqualifies participants from purchasing guns. But since most suicides involve guns, it is possible the effect on total suicide is driven by gun suicide alone.


Introduction
In 1996, California became the first state to legalize medical marijuana. Known officially as the Compassionate Use Act, Proposition 215 allowed patients and caregivers to cultivate and possess marijuana for medical use. The campaign in favor of Proposition 215 focused on the benefits for seriously ill patients. Claiming that the Proposition "sends our children the false message that marijuana is safe and healthy," the campaign against the Proposition focused on anti-drug education (Legislative Analyst's Office, 1996, p. 4). Neither side addressed potential public health sequelae.
If Proposition 215 led to an increase in marijuana use, for example, might it also lead to higher rates of all injury deaths (Gerberich et al., 2003) including deaths from assault (Resnick et al., 2004), deaths from motor vehicle crashes (Li et al., 2012) and, the subject of the present study, deaths from suicide (Petronis et al., 1990)?
These sequelae assume that medical and recreational users are similar. With one exception, the evidence supports this assumption. Since most California medical users were introduced to marijuana as recreational users, for example, it is reasonable to assume that the user-types have similar socio-economic backgrounds (Grella et al., 2014). Compared to recreational users, however, California's medical users were more likely to report early health problems or disabilities that would warrant medical use (Lankenauer et al., 2018). Although Proposition 215 was drafted so loosely that it effectively legalized all uses of marijuana (Vitiello, 1998), marijuana use by California juveniles, who were not eligible for medical marijuana certificates, did not increase following Proposition 215 (Khatapoush and Hallfors, 2004).
Electronic copy available at: https://ssrn.com/abstract=3391891 Nevertheless, at the national level, during a 15-year period when a majority of states loosened their control of medical marijuana, the U.S. suicide rate rose by 24 percent (Curtin, Warner and Hedegaard, 2016), prompting many to question how legalization and suicide may be linked. The systematic evidence connecting this trend to the availability of medical marijuana is ambiguous, however. Rylander, Valdez, and Nussbaum (2014), for example, find no correlation between a state's suicide rate and the number of medical marijuana cardholders in the state. Similarly, comparing suicide before and after a state enacts a medical marijuana law,  find no change in a state's suicide rate. In contrast, Anderson, Rees and Sabia (2014) report a 10.8 percent reduction in suicides averaged across all medical marijuana states.
Attributing a suicide trend to the availability of medical marijuana raises questions about the potential mechanisms at play. What theoretical mechanisms could lead us to expect a relationship between the availability of medical marijuana and suicide? Could these mechanisms be more salient for certain types of suicides than others? If the expected relationship is observed, what methodological rules could be used to support a causal interpretation of the relationship? We address these questions in order.

Etiology of Suicide
Sociological theories of suicide follow Durkheim's (1951Durkheim's ( [1897) dictum that "suicide varies inversely with the degree of integration of the social groups of which the individual forms a part" (1951, p. 209). Institutions that successfully integrate the individual, providing a sense of belonging to the community, inhibit suicide. When institutions break down, the community ties that might otherwise inhibit suicide atrophy, Electronic copy available at: https://ssrn.com/abstract=3391891 and suicide increases. Durkheim used cross-sectional correlations between suicide and the strength of religious, familial and socio-economic institutions to demonstrate his theory. The theory has been used to investigate relationships between suicide and unemployment, (Almgren et al., 1998;Coope et al. 2014;Phillips and Nugent 2014), poverty and income inequality (Burr et al. 1999;Curtis et al. 2013;Kubrin et al., 2006), divorce and family structure (Baller and Richardson, 2002;Stockard and O'Brien 2002;Sun and Zhang 2016), immigration and cultural assimilation (Wadsworth and Kubrin, 2007), and cohort size (Stockard and O'Brien, 2002). Regardless of focus, research largely advances a motivational argument for understanding variation in suicide rates across place or time.
Although legalization of medical marijuana is likely to affect a range of societal institutions, the indirect effects through these institutions are expected to accrue gradually. Individual-level direct effects, in contrast, are expected to be realized abruptly. A more appropriate individual-level theory for explaining the relationship between medical marijuana and suicide posits suicide risk as the product of motivation and opportunity factors. That is,

Risk = Motivation x Opportunity
Holding motivation constant, suicide risk responds to changes in opportunity. Holding opportunity constant, risk responds to changes in motivation. Chew and McCleary (1994) use motivation-opportunity mechanisms to explain life-course changes in suicide. Kubrin and Wadsworth (2009)

use motivation-opportunity mechanisms to
Electronic copy available at: https://ssrn.com/abstract=3391891 explain the effects of socio-economic factors and firearms availability on race-specific suicide. Wadsworth, Kubrin and Herting (2014) use motivation-opportunity mechanisms to explain suicide trends for young Black males. Consistent with this literature, we argue that, if medical marijuana affects suicide risk, it must do so through one or both pathways.
Mental health theories operate through a motivation pathway. The psychiatric consensus is that suicide is related to depression, anxiety and other treatable disorders, for example (Mann et al., 2005;Turecki and Brent, 2016;Zalsman et al., 2016). If marijuana alleviates the acute stress associated with these disorders, we expect suicide risk to decrease following legalization of medical marijuana. The evidence for this is mixed, however (Anderson, Rees, & Sabia, 2014;Rylander Valdez, & Nussbaum, 2014;Moore et al., 2007;Calabria et al., 2010). Whereas medical users report that alleviation of acute symptoms of these disorders was a primary motivation for permit applications (Grella et al., 2014;Greer et al., 2014), a systematic review by Walsh et al. (2017) reported that this was not consistently observed across credible studies. Of course, marijuana use itself may constitute a risk factor for suicide apart from alleviating symptoms related to depression and anxiety, at least among some populations and for some levels of suicidality (Stack 2014).
The relationship between alcohol and suicide also operates through a motivation pathway. A meta-analysis by Darvishi et al. (2015, p. 1) supports the strong consensus that alcohol use disorder "significantly increases the risk of suicidal ideation, suicide attempt, and completed suicide." With respect to medical marijuana, of course, the theoretical prediction depends on whether marijuana is used in addition to or instead of Electronic copy available at: https://ssrn.com/abstract=3391891 alcohol. If marijuana and alcohol use are combined, one might expect no change in suicide risk or even an increase in suicide following legalization. If marijuana replaces alcohol, on the other hand, one might expect a decrease in suicide risk following legalization. Self-reports by medical users in California (Reiman, 2009) and Canada (Lucas et al., 2013) indicate that a substantial proportion substitute marijuana for alcohol and other drugs. Since the instruments included items about patients' use of potentially criminal behaviors, self-reports are potentially biased.
With a few exceptions, opportunity pathways have received less attention in the suicide literature. Noting that suicide risk is highest when the victim is alone in the absence of guardians who would otherwise intervene (Wassernan and Stack 2008, p. 759), Chew and McCleary use the routine activity theory (Cohen and Felson, 1979) to explain why risk is relatively lower on weekend days, when other household members are more likely to be present, and relatively higher on Mondays, when other household members are likely to be out of the home at school or work. By analogy, if access to medical marijuana obviates the need to leave home, one might expect a lower suicide risk following legalization. If medical users substitute marijuana for alcohol, moreover, legalization may result in less time spent in licensed alcohol establishments. Anderson, Hansen and Rees (2013) use this argument to explain their finding of a reduction in motor vehicle fatalities following legalization.
Firearms access is a relevant opportunity pathway. The positive correlation between firearms access and suicide risk has been demonstrated for U.S. metropolitan areas and counties (Miller et al., 2013). These correlations are subject to the ecological fallacy, however. At a disaggregated level, compared to matched controls who live in Electronic copy available at: https://ssrn.com/abstract=3391891 non-gun households, individuals who live in gun households are 3.4 times more likely to die of suicide. Comparable risks have been reported for subpopulations, including middle-age and older adults (Conwell et al., 2002), adolescents (Brent et al., 1991) and women (Bailey et al., 1997).
The 1993 Brady Handgun Violence Prevention Act prohibits the purchase of guns by individuals who are addicted to controlled substances. Though used for legal medical purposes, marijuana remains a controlled substance under U.S. law. 1 Since California medical marijuana users were not allowed to purchase firearms in 1997, and since California firearms regulations are relatively strict, we expect a reduction in suicide risk following legalization.
In sum, in 1996, California legalized marijuana use for medical purposes.
Implementation was abrupt and uniform, presenting a natural experiment that we take advantage of in order to estimate the causal effect of a medical marijuana initiative on suicide risk. In the current study, we aggregate total, gun and non-gun suicides by state for the years 1970-2004. Using a synthetic control group quasi-experimental design (Abadie, Dimond and Hainmueller, 2010;2015), we construct a control unit for California from time series of the 41 states that did not legalize marijuana during the time-frame. We interpret post-intervention differences for California and its synthetic control time-series as the effects of the medical marijuana law on suicide. Significance of the effects are assessed with permutation tests.

Data and Methods
Ideal counterfactual control units seldom exist in nature. However, given an ensemble of less-than-ideal untreated units, an ideal counterfactual can often be approximated as a weighted sum of the untreated units. In order to create Counterfactual California, our control unit, suicide death certificates from the National 3 Prior to the ICD-8a revision adopted in 1968, physicians did not report whether external causes of death were intentionally self-inflicted, accidental, or malicious. Under the ICD-7, a firearm suicide would have been coded as a general explosive fatality, with no indication of intentional self-harm, accidental injury, or intentional suicide (Gjertsen, 2015). time series were scaled as proportionate changes from their 1968 values.
With the time frame and donor pool defined, if a suitable synthetic control state exists, it can be constructed from a constrained factor analysis of predictors (Abadie, Dimond and Hainmueller, 2010;2015). To introduce this method, suppose that an intervention breaks a time series of T observations, Y 0,t , into pre-and post-intervention segments of T0 and T-T0 years respectively. If Y j,t is the t th pre-intervention year (t=1,…,T0) for the j th donor state (j=1,…,J), the corresponding observation of the synthetic control state is the sum, The w1, ... , w J weights are selected under two constraints. Since the internal sum is equivalent to Y syn,t , minimizing the RMSPE statistic optimizes the "fit" of between the affected time series, Y 0,t , and the corresponding synthetic control time series.
Because Y syn,t is unobserved, the optimization relies on a factor analysis of variables that predict Y 0,t . Abadie and Gardeazabal (2003) describe two approaches to the optimization. The first approach uses known causal predictors of Y 0,t . This method necessarily assumes that the causes of Y 0,t are known, measured, available, and stable across units and time. If the causes are unknown, unmeasured or unavailable, the observed values of Y 0,t can be substituted. McCleary, McDowall and Bartos (2017) extend this second approach by including first-differences. The causal predictors of suicide are unknown, unmeasured or unavailable, so we follow this data-driven approach.
To evaluate the significance of the 1996 intervention on the suicide outcomes, we adopt the permutation inference framework proposed by Abadie and Gardeazabal (2003) and extended by Abadie, Diamond and Hainmueller (2010). By iteratively reassigning the treatment condition to untreated donor pool units, this "in-sample placebo test" generates a distribution of observed placebo effects among the untreated donor pool units. If the estimated effect for California is larger than the effector for any of the untreated donor pool states, we can infer that the estimated impact of California's 1996 intervention is significant, thereby rendering spuriousness implausible. Figures 1-3 plot California's total, gun and non-gun suicide rates against their synthetic control states. Visual comparisons of the pre-intervention time series reinforces our interpretation of the RMSPE statistics. Post-intervention differences are more difficult to interpret. While it appears that rates of total suicides and gun suicides drop significantly in the aftermath of Proposition 215, the impact of medical marijuana on rates of non-gun suicides are ambiguous.

Results
If treatment and control units are equivalent prior to the intervention, the causal validity of a synthetic control inference can be justified with a difference-in-difference rationale. To estimate the unemployment effects of an increase in the minimum wage, for example, Card and Krueger (2004) observe 400 fast food stores in New Jersey and Pennsylvania before and after a minimum wage increase in New Jersey. Because the stores were made equivalent prior to the intervention, and because the minimum wage rose in New Jersey but not in Pennsylvania, the difference in before-after differences could be interpreted causally. Table 1 reports the non-zero wj weights for synthetic total suicides, synthetic gun suicides and synthetic non-gun suicides. The values of the wj weights are interpreted relative to the set of non-zero weights. In this instance, 23 of J=41 states have weights of zero and, hence, contribute nothing to the constructions. The non-zero weights for Connecticut, the District of Columbia and Ohio contribute to all three synthetic controls;

Figures 1-3 Here
Kentucky and Texas contribute to two of the three synthetic controls; Alabama, Idaho, Illinois, Kansas, Louisiana, Massachusetts, North Dakota, New Hampshire, New Jersey, Pennsylvania, South Carolina, Wisconsin and Wyoming make unique contributions to a single synthetic control. Contributions range from very small (w j =0.013) to moderately large (w j =0.324). No state makes an overwhelming contribution to any of the synthetic control states. The gun suicide synthetic control "fits" best (RMSPE=0.02597), the nongun suicide synthetic control "fits" worst (RMSPE=0.06219) statistics. The "fit" of the total suicide synthetic control falls midway (RMSPE=0.03125) between the best and worst.

Table 1 Here
In this particular instance, California t and Synthetic California t follow the same path prior to the 1997 intervention, 1970, …, 1996 Beginning in 1997, however, the expected difference is the quantity α t . E(California t -Synthetic California t ) = α t for t = 1997, …, 2004 The quantity α t is not necessarily zero nor constant for different values of t. Allowing for approximate equivalence, estimates of α t will be tested for significance.

Table 2 Here
The first two columns of Table 2  Results from the in-sample placebo tests for all suicides, gun suicides, and nongun suicides are reported in the last column of Table 2. For all suicides, California's mean post-intervention effect was 4.6 times greater than its pre-intervention RMSPE, ranking second among 41 donor pool states (p<.05). For gun suicides, California's mean post-intervention effect was 6.8 times greater than its pre-intervention RMSPE, ranking first among 41 donor pool states (p<.05). In contrast, for non-gun suicides, California's mean post-intervention effect was only 1.7 times greater than its preintervention RMSPE, ranking 20 th among 41 donor pool states (p≥.488). While the insample placebo tests for all suicides and gun suicides suggest that the estimated impacts are unlikely to reflect spurious artifacts and are larger than the idiosyncratic shocks observed among the untreated donor states, the in-sample placebo test for nongun suicides cannot be interpreted with the same confidence. Since gun suicides and non-gun suicides are the two nested subtypes of suicide, the MML effect for all suicides appears to be driven by the reduction in gun suicides.

Discussion
In 1996, California voters passed an initiative Proposition 215, which legalized marijuana use for medical purposes. Because the Proposition was implemented in an abrupt and uniform manner, legalization presented a "natural experiment." To estimate the causal impact of legalization on suicide, annual time series of total, gun and non- Findings also reveal, however, that legalization's impact on non-gun suicides is considerably smaller and arguably not different than what would be expected to occur by chance. Confidence in these findings is underscored by the methodological approach undertaken in the study. A strength of the Synthetic Control Group Design is that it allows us to examine the net effect of medical marijuana legalization on suicide.
Despite the strengths of this design, important limitations remain, many of which present opportunities for future directions in research. Because we examine suicide trends over eight post-intervention years, we are fairly confident that the effects are permanent. Because our time series end in 2005, on the other hand, it is difficult to generalize our theoretical result to subsequent years. We are limited by the fact that medical marijuana laws began to proliferate across the U.S. after 2005, threatening to contaminate the "donor pool" of untreated states In virtually all of the states that legalized medical marijuana after 2005, moreover, reforms were not implemented abruptly or uniformly, thereby not lending themselves to simple, confident causal interpretations.
Another limitation which presents a future direction relates to the mechanisms that may account for the findings of the study. What are the mechanisms responsible for the sharp decline in total, but especially, gun suicides following medical marijuana legalization in California? We proposed mechanisms related to the substitution of marijuana for alcohol and other related substances; marijuana use itself which may reduce actual motivation for suicide; the inability of medical marijuana patients to purchase firearms; and, changes in the culture of recreational substance use leading to fewer unsupervised opportunities to commit suicide in the home. Each of these pathways should be tested, although many will require the collection of additional data to do so. For example, one likely fruitful research direction would be to collect annual data on alcohol consumption in California and assess whether it is a plausible mechanism by which medical marijuana legalization could cause a reduction in gun suicides. Beyond adjudicating these various pathways, testing mechanisms may reveal insight into why we do not find the expected reduction in non-gun suicides following legalization. Unfortunately, we do not have the data to test these mechanisms, yet it will be essential for future research to do so.  Electronic copy available at: https://ssrn.com/abstract=3391891