International Review of Law and Economics 37 (2014) 180188

Contents lists available at ScienceDirect

International Review of Law and Economics

Firearms and suicides in US states

Justin Thomas Briggs, Alexander Tabarrok
Department of Economics, George Mason University, Fairfax, VA, United States

a r t i c l e

i n f o

a b s t r a c t
This study investigates the relationship between rearm prevalence and suicide in a sample of all US states over the years 20002009. We nd strong, positive effects of gun prevalence on suicide using OLS estimation, across a variety of measures for gun possession, and with several sets of controls. When using instrumental variable estimation, the effect remains signicant, despite also nding signicant evidence that gun ownership causes substitution towards gun-suicide rather than other methods of suicide. There is also evidence for non-linearities in the effects of guns on suicide. 2013 Elsevier Inc. All rights reserved.

Article history: Received 17 June 2013 Received in revised form 28 September 2013 Accepted 6 October 2013 JEL classication: I12 I19 K10 Keywords: Firearms Suicide Guns

1. Introduction Firearms play a unique role in public discourse. The US Constitution protects the right to bear arms. For some, this right represents an important safeguard against tyranny. For hunters and sportsmen, rearms enable a vibrant recreation. Firearms also play an important but largely unknown role in self-defence. Yet in 2010, the latest year for which there are complete gures, there were 19,392 suicides, 11,078 homicides, and 606 accidental deaths by rearm, in addition to 73,505 non-fatal injuries by rearms (Centers for Disease Control and Prevention, 2012). Unfortunately, even basic information such as how many households own rearms is irregular and partly as a result there is little scientic consensus on how rearms inuence violent injuries. Although the effect of rearms on homicides has been a topic of recurring debate, less attention is often given to suicide, despite there being more deaths attributable to suicide than to homicide. This may be in part because people view suicide as a private decision only affecting friends and family of the deceased, although this impact should not be minimized. But many psychological studies nd that suicides are frequently impulsive decisions (e.g. Simon et al., 2002), and that less than 10% of suicide survivors go on to successfully re-attempt suicide over the long term (Owens,

Horrocks, & House, 2002). Few suicides appear to be considered choices. In this study we specically explore the relationship between rearm ownership rates and rates of suicide, using a newly constructed dataset covering US states from 2000 to 2009. We utilize all data from the rst state-level representative survey of gun ownership, as well as four other proxies thereof, including one new to the literature. In addition, we will develop instruments for rearm ownership rates. 2. Background The United States stands out among developed countries for having the most rearms (Geneva Graduate Institute of International Studies, 2007). The US also has the highest rate of suicide by rearm of any developed country. Indeed, the rate of suicide is on par with the rate of trafc deaths in 2009, there were 11.0 trafc deaths per 100,000, and 12.0 suicides and the majority of suicides are by rearm. However, the total suicide rate is also lower than in many countries with fewer rearms. More generally, crosscountry studies tend to show no relationship between rearms and suicide (Killias, van Kesteren, & Rindlisbacher, 2001; Krug, Powell, & Dahlberg, 1998). Within the US, ecological and case-study approaches both appear to show a strong positive association between rearms and suicides by rearm (e.g. Azrael, Cook, & Miller, 2004; Brent et al., 1994; Kleck & Patterson, 1993). What is less clear is whether there is an association between rearm ownership and overall

Corresponding author. E-mail addresses: [email protected] (J.T. Briggs), [email protected] (A. Tabarrok). 0144-8188/$ see front matter 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.irle.2013.10.004

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188

181

suicide. After all, in the absence of a rearm, there are several other methods of suicide, most predominately suffocation and poisoning, and a determined individual could simply substitute the means to achieve the same end. For example, Sloan, Rivara, Reay, Ferris, and Kellermann (1990) compare Vancouver and Seattle and nd that Seattle has more rearms and suicides by rearm than Vancouver but it has more overall suicides in only the 1524 age group which suggests a signicant degree of method substitution. On the other hand, policy change studies have shown that decreased access to rearms can decrease suicide overall. In one careful study, Leigh and Neill (2010) nd that the mandatory buyback of 20% of the stock of rearms in Australia in 1997 (which halved the number of households holding rearms) led to nearly an 80% reduction in suicides. Similarly, Lubin et al. (2010) show that when drafted soldiers in the Israeli military were prohibited from bringing their guns home with them for the weekend, suicide fell by 40%, owing to a fall in suicides over weekends that was not accompanied by an increase in weekday suicides. Is there a causal relationship between rearm ownership and suicides? In this paper we will focus on how the total numbers of guns in the hands of citizens affects suicide rates on the margin. There are three potential causal pathways by which increased rearm ownership may increase successful suicide. First is the weapon instrumentality effect, which posits that if guns are inherently more lethal than other methods then they would make attempts more likely to be successful. In 2001, 85% of suicide attempts by gun were lethal, compared with 69% for the next most lethal method (suffocation/inhalation) (Vyrostek, Annest, & Ryan, 2004). Intent, however, is a confounding variable. One would suspect that those more intent on suicide would choose more lethal methods, although the evidence for this proposition is weak (compare Peterson, Peterson, OShanick, & Swann, 1985 with Hamdi, Amin, & Mattar, 1991). Second is the method selection (or induction) hypothesis, positing that the actual or perceived ease of committing suicide by gun may make it more likely to happen (because it is thought to be either quicker, simpler, or more likely to be successful). Psychologists have shown that suicidal thoughts are often only brief, and many suicide attempts are impulsive (Hawton, 2007; Simon et al., 2002), so the absence of an effective immediate means may prove a deterrent. Interviews with people who failed at suicide often indicate that methods were chosen according to ease of availability (Peterson et al., 1985; Skopek & Perkins, 1998). Third, if it is an easier, quicker, and more certain means of selfdestruction than others, it may reduce the chances of successful outside intervention. The only plausible source of reverse causality (as opposed to correlation) is that an individual may purchase a rearm for the purpose of committing suicide.1 But even if all rearm suicides represented a new gun purchase, this would account for only 0.2% of rearm purchases2 so this does not appear to be a potent form of reverse causality. Firearms may also be related to suicide via other unobserved avenues that lead to correlation without causation, such as prevalence of generalized mistrust, mental health issues, or morbid thoughts. Or there may be no causal relationship at all. 3. Data We will analyze state-level rates of suicide, considering both rearm- and non-rearm-varieties as well as overall rates, in several samples covering the decade 20002009. The mortality data

14 12 10 8 6 4 2 0 1985 1990 1995 2000 2005 2010

Firearm suicides

Non-rearm suicides

Fig. 1. Trends in suicide rates 19852010 (per 100,000).

cover substantially all such deaths in the US, and are from Centers for Disease Control and Preventions (2012) WISQARS compressed mortality le. For context, we provide national trends in suicides in Fig. 1. Suicides rose slightly over the study period, particularly in the category of non-rearm suicides. Summary statistics for all data are in Table A.1. 3.1. Firearm measures The variable of interest is the percentage of individuals living in households with at least one rearm.3 A major shortcoming of much earlier work is that no data was available that directly measured gun ownership in a representative way in all states. Most states do not require registration or licensing, so consulting administrative records is not an option (Vernick & Hepburn, 2003). In this paper we use data from the rst and only national survey that is representative at the state level. The survey data was collected in the large Behavioral Risk Factor Surveillance System (Centers for Disease Control and Prevention, 2010) telephone survey conducted by the states under the guidance of the CDC. The question Are any rearms kept in or around your home? (or a slight variation thereof) was asked in 2001, 2002, and 2004. After 2004, the question was discontinued. Respondents who refuse to answer, or do not know, together accounting for 4 percent of the total or less, are excluded.4 Although we believe the BRFSS measure is the best survey measure yet to be used, it still remains awed. We therefore test for robustness using a variety of other proxies. Namely, rates of Google searches for gun-related terms, the proportion of suicides that are committed with a gun, the rate of background checks for gun purchases and the rate of accidental death by rearm. The latter three proxies we combine into a gun composite index using their rst principal component.

1 Wintemute, Parham, Beaumont, Wright, and Drake (1999) present evidence that at least some people do purchase rearms for the purpose of committing suicide. 2 In 2009 e.g., there were 18,735 rearm suicides and 8,968,180 net new rearms.

3 The literature implicitly assumes that household ownership means enhanced access to a rearm although this assumption requires that there is some barrier to buying or obtaining a rearm in order to commit suicide. There are several reasons to think that although by no means insurmountable, the need to nd and purchase a gun may be enough to deter in many cases. First, it takes time and effort to nd and go to a rearm dealer, choose a weapon, get a background check, and obtain it (possibly after a mandatory waiting period), and this may give time for reection and re-evaluation. Second, for many the price of a rearm and ammunition is not trivial, starting at several hundred dollars. Third, a potentially suicidal individual may be paralyzed by oppressive thoughts that preclude seeking to purchase a rearm. Finally, there may be cognitive barriers that cause an individual to think obtaining and properly using one would be harder than it is. For there are to be some effect of current possession rates, any marginal barrier will sufce. Ownership is likely also correlated with other measures of ease of access. 4 California and Hawaii each had a missing observation that was estimated using the other observations and national trends.

182

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188

We use rates of Google searches containing any of the following keywords: gun, rie, pistol, handgun (Google, 2012). Available beginning in 2004, these rates represent the relative frequencies of these search terms within the state, compared with other searches that year. Used for the rst time here, this Google Composite Gun Searches measure proxies for gun possession under the assumption that interest in guns is proportional to household ownership rates, and rates of Google searches for these terms accurately measure interest.5 This is not implausible, as several studies do nd Google Searches to be useful indicators of consumer sentiment (Choi & Varian, 2012; Della Penna & Huang, 2010), and as we show below, this index correlates well with the BRFSS measure. Next, we use a commonly cited proxy for gun ownership the proportion of suicides that are committed with a gun. This is constructed from CDC data, has been shown to have high crosssectional correlation with survey measures of ownership (Azrael et al., 2004), and is available for all years 20002009. Although this is commonly used, there has been some question about how accurate it is (Duggan, 2001; National Academy of Sciences, National Research Council, 2004; Shenassa, Daskalakis, & Buka, 2006), with e.g., the latter nding it not strongly related to self-reported neighborhood level of rearm carrying and availability among juveniles. The rate of background checks for gun purchases for a given state is available for the full sample, 20002009, from the Federal Bureau of Investigations (2012) National Instant Criminal Background Check System (NICS). If background checks measure gun purchases and purchases are proportional to existing ownership rates (such as if rearms depreciate, become obsolete, and are replaced at a certain rate), then they would be a good proxy for rearm ownership rates. Unfortunately the concerns about background checks are well known: they are only conducted by federally licensed rearm dealers, and many purchases are conducted outside these dealers; a background check may represent zero, one, or many gun purchases6 ; and due to varying state regulations, raw background checks may not be comparable over time or between states. For example, Kentucky began requiring background checks for concealed carry permits in 2006, greatly inating numbers. To the extent possible, these internal comparability issues have been corrected in our data. Nonetheless, the data remain imperfect. The rate of accidental death by rearm would perhaps be one of the most defensible proxies, except for the fact that it is fairly rare and as a Poisson random variable with low expected value, it likely suffers truncation error. In an attempt to overcome weaknesses and combine the strengths of some proxies we have constructed a composite measure of rearm ownership (Gun Possession Composite) from the rst principal component of these three proxies: percent Suicides By Gun, background check rates, and rates of accidental death by gun. This captures over 74% of the total variation of the component variables, all of which contribute positively, and it correlates strongly ( = 0.84) with the BRFSS ownership measure. In Table 1, we provide a comparison of some salient features of the gun measures used in this paper. As the chart makes clear, no one measure is ideal. By using a wide variety, we can have enhanced condence if and when they yield similar results. Over the study period, at a national level, all rearm measures except

that of background checks show a slight downward decline.7 Background checks (consistent with net new rearms) increased over the period, even after accounting for the increase in the number of households, suggesting that proportionally fewer households are holding more rearms each. All measures show substantial variation across states. Crosssectionally, the four measures of gun possession we consider are highly correlated, and BRFSS is correlated with all other proxies at > 0.8. Scatter plots are shown in Fig. 2 for years in which there is an overlap. Finally, we use Cronbachs alpha to evaluate whether these appear to measure the same construct, nding = 0.84, which denotes good internal consistency, giving increased condence that these measures do in fact represent the latent gun possession variable. 3.2. Other controls We use a number of controls common in the literature. Our baseline controls include the population (US Census Bureau, 2012b), poverty rate (% under the poverty line, all ages) (US Census Bureau, 2012d), annual average unemployment rate (Bureau of Labor Statistics, 2012), percent urban land area (Demographia, 2012), percent urban population (US Census Bureau, 2012a), Gini coefcient of household income inequality, calculated from US Census Bureau (2012b) data, prevalence of drug and/or alcohol abuse or dependence in the population aged 12 (Substance Abuse and Mental Health Services Administration, 2012), and prevalence of frequent mental distress (FMD)8 among non-institutionalized adults (Centers for Disease Control and Prevention, 2010). Additionally, we use demographic controls for groups at high risk for suicide, including the percent of males age 65 and higher, and the percent white. We use time-specic xed effects to control for any time-varying factors that affect all states. Finally, we control for regional variation using census region xed effects. If for example the southeast region were culturally more predisposed to gun violence (Nisbett & Cohen, 1996; Rosenfeld, Messner, & Baumer, 2001), this should capture that effect. As a baseline, we believe these controls represent a good set of outside factors affecting suicide, are likely exogenous, and should not omit too much. Further controls are explored later in the paper. 4. Approach and results We attempt to ascertain the association between rearm possession rates and intentionally self-inicted fatal injuries. As discussed above, this is fraught with difculty because of the variety of confounding inuences and because of difculties in measuring possession rates. We attempt to improve our knowledge about this association by using a multifaceted approach. First, to evaluate sensitivity of these results to omitted variables and endogeneity biases, we use a variety of controls for potential drivers of suicides, ranging from a very minimal specication to a very complete set of controls. Second, we use a pair of instruments for gun possession, in an attempt to tease out causality. Finally, as mentioned, we explore four measures of gun possession. The basic model is of the following form: ln Suicidesit = 0 + 1 GunPossessionRateit + X it + Ri + Tt +
it

This measures the state-specic variation in gun searches, averaged over a year. As such many outside drivers of interest in guns, such as national gun-control debates or incidences of widely covered shootings, would not be included. 6 In practice, one is almost certainly the modal number of guns purchased per adjusted background check. In addition we have been able to conrm that at a national level over the 10 year study period, the Pearson correlation coefcient between NICS checks and net new rearms in the US (domestic rearm production plus net rearm imports) (Bureau of Alcohol, Tobacco, Firearms and Explosives, 2011) is 0.78, signicant at the 1% level.

7 National polls such as the general social survey (GSS) (Smith, Marsden, Hout, & Kim, 2011) and Gallup polls (Gallup, 2011) also show declines over the period. 8 FMD is measured by counting those answering 14 or more days to the question: Thinking about your mental health, which includes stress, depression, and problems with emotions, for how many days during the past 30 days was your mental health not good?. One state-year observation was missing and was estimated.

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188 Table 1 Characteristics of gun measures. Measure Govt health survey X Samples non-instl phone users X X X X X X X X X X X Internet users of Google search only Sometimes sparse Behavioral proxy Requires proportionality assumption May capture purpose-bought guns

183

BRFSS Google Gun Searches % of Suicides By Gun Acc. deaths by gun Background check

Firearms in Household (BRFSS survey)

80 60 40 20 80 60 40 20 60
AL LA MS WY AR SC WV KY ID TN MT GA AK OK NC AZ VA VT TX MO IN NV NM ND OR KS NE WA PA ME OH CO UT FL MI MN DE SD MD IA WI CA NH IL CT NY NJ RI MA HI AK WY MT WV OK ID AR ND MS SD AL NM TX TN MI KY KS NE NV AZ OR MO ME UT SC OH CO WI LA MN IN VA WA IA NC VT FL NHPAGA CA IL CT IL CT DE RI RI MD NJ MA NJNY MA NY HI HI KY AL MS WV MT AR LA TN WY ID SD SC AK OK NM OR VT CO MO GA AZ PA NC KS ND VA TX IN DE NV NE WI OH ME NH UT MN FL WA MI IA CA MD IL IL CT CT NY NJNY NJ RI RI MA MA HI HI

Percent of Suicides By Gun

AK WY MT WV OK ID AR MS SD ND AL NM TX MI KY KS NE NV AZ TN MO OR ME UT SC OH CO WI LA MN IN VA WA IA NC PA VT GA FL NH CA MD DE AL KY MS WV MT AR LA TN WY ID SC AK OK NM OR VT CO MO GA ND AZ PA NC KS VA TX IN DE ME NV NE WI OH NH UT WA MN MI FL IA CA MD SD

Gun Composite Google Searches

AL KY AR WV MT LA TN MS WY SDID SC AK NM OK OR VT CO MO GA ND AZ PA NC KS VA TX IN DE ME NV NE WI OH NHWA UT MN FL IA MI CA MD IL CT NY NJ RI MA HI

40

Gun Possession Composite (PCA)

20 0

20

40

60 20

40

60

80 20

40

60

80

Fig. 2. Scatter plots of gun possession measures. Note: All scatter plots are of a three year average (2001, 2002, 2004) to coincide with the BRFSS sample years, except the GunCompSearches vs. BRFSS scatter plot, where only 2004 overlaps.

We wish to evaluate 1 , the effect of gun possession rates on Suicides (the number of suicides, or some subset thereof),9 across US states i and years t, after controlling for a set of time and statevarying controls Xit as discussed elsewhere, always including xed effects for year (Tt ) and region (Ri ), as well as log population. Standard errors are clustered by state in all regressions. We do not use state xed effects because of the durability of rearms, and because the rates of rearm possession have little actual year to year variance over the study period. The section proceeds by reviewing OLS regression results using the BRFSS survey measure of gun possession. Then we review the use of varying sets of controls in this regression, before moving on to consider other measures of gun possession in a similar fashion. Next, we consider the possibility nonlinear effects of gun possession. Finally, we use an instrumental variable approach on the same specications. In all cases, we run three sets of regressions: First on the gun suicide rate, then the non-gun suicide rate and nally the overall level of suicides. The overall rate is perhaps the most important rate but this is also the most complicated to interpret and the two earlier

regressions will aid in interpretation. If guns are to have an effect at all, we expect this to appear at a minimum in the gun-suicide regressions. Guns should affect gun suicides most of all, for example, and if guns affect suicides overall, it should be through this rate. We investigate the non-gun suicide rate primarily to assess whether there is substitution. For example, if people choose to use a gun simply because its convenient, but would have performed the same act even without a gun, we should see a negative coefcient on non-gun suicide rates. We may also see evidence of a third causal variable across the regressions. If, for example, increased social anomie increases gun purchases and all types of suicides we should see a positive correlation between gun possession and nongun suicide. Careful examination of all three regressions will help in interpretation. 4.1. BRFSS The most direct and possibly the most representative measure of gun possession rates in US states comes from the BRFSS survey conducted in 2001, 2002, and 2004. We begin by analyzing suicide with this data. To get a quick sense, rst see Fig. 3. The left-hand plot shows the simple relationship between gun possession and suicide. The right hand side shows the same relationship, after controlling for other differences between states using our baseline controls

9 We refer to regressions on the suicide rate later because of the interpretation of the regression coefcients as a semi-elasticity.

184

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188

.00015

AK MT WY

.00004

NV

AK WY PA

NV AZ CO

NM OK TN KY VT LA UT SC MO

IDWV AR MS AL ND

e( gunSuicideRate | X ) .00002 0 .00002

Rate of Suicides by Firearm .00005 .0001

OH IN NM VT IL NC

AZ TN

FL MO KS NJ OKMT GA VA

CT LA

OR FL DE MD CA IL CT HI NJRI MA NY WA NC GA ME INKS MI

VA TX OH PA NH

SD

NE WI IA MN

ME

AL MI NY IDAR SC CO NE ND MS MD WV SD MN WI DE RI TX NH MA KY UT OR IA HI WA

CA

10

20

30 40 Firearm in household (BRFSS)

50

60

.00004

20

10

0 e( guninhouse | X )

10

Fig. 3. Scatter plots of BRFSS gun possession vs. suicide. Note: All scatter plots use a three year average (2001, 2002, 2004) to coincide with the BRFSS sample years. The plots show gun possession rates on the vertical axis and rearm suicide rates on the horizontal. The left-hand plot is a simple relationship. The right-hand plot is conditional upon the baseline controls (in these charts we are viewing rates for visual clarity; following regression analyses consider log counts while also controlling for log population).

(discussed above). There is a clear positive association between gun possession and suicide in both versions. 4.1.1. Basic OLS approach More formal results are shown in Table 2. Following the model above, we use the baseline set of controls for three sets of regressions on suicide. In column (1), the dependent variable is gun suicide, in column (2) it is non-gun suicide and in column (3) all suicides. The BRFSS gun possession variable (guninhouse) is the ) shows that independent variable of interest. The OLS estimate ( 1 a one percentage point rise in the prevalence of individuals having rearms in their household is associated with an expected mean increase of 1.7% in the number of gun suicides (signicant at the 1% level).10 In the results for non-rearm suicides in column (2), we do see some sign of method substitution for suicide a 1% rise in gun prevalence is associated with a 0.8% decline in non-gun suicides but the amount of substitution is not enough to mitigate an overall rise. As shown in column (3) a 1% increase in gun prevalence is associated with an overall 0.5% rise in the number of suicides (signicant at the 10% level).11 4.1.2. Varying specications Next we investigate the robustness of these results to varying the set of controls used. In addition to the baseline specication used above, we will use a minimal and full specication. The minimal specication is a subset of baseline, including only demographic controls, percent urban land & population, and unemployment & poverty rates in addition to population and year & regional xed effects. These controls were selected for being as independent as possible from proposed causal mechanisms for suicide, while still capturing much of the diversity between states. It may suffer from omitted variables bias but is unlikely to have serious endogeneity issues.

The full specication is a superset of the baseline. In addition to the baseline controls, it includes median household income (US Census Bureau, 2012d), percent of children living in a single mother family, percent of divorced adults (US Census Bureau, 2012b), and distance in miles to the nearest hospital emergency room,12 which is meant to proxy for access to mortality-reducing health care (Health Forum and ESRI, 2008; US Census Bureau, 2012c), and a measure of social connectedness,13 which is often thought of as a factor in preventing suicide (Corporation for National & Community Service, 2012). The full specication has the reverse strengths and weaknesses of the minimal: while it is fairly unlikely to suffer from omitted variables bias, it is more likely to suffer from endogeneity issues. Results are presented in columns (4)(9) of Table 2. Broadly, they are similar to the baseline specication for both the minimal and full specications. The minimal specication gives effectively the same results as the baseline across gun, non-gun, and overall suicide. The full specication shows a slightly weaker association with gun suicide, with one percentage point increase in gun prevalence implying a 1.4% increase in gun suicides, and a slightly stronger substitution effect, the same 1% increase in rearm ownership implying a 1.1% fall in non-gun suicides. As a result, the effect of a one percentage point increase in gun ownership rates on suicide rates is reduced from 0.5% to a no longer statistically signicant 0.3%.

4.1.3. Divorce and distance to hospital Among the most interesting ndings on the control variables are that the percent of divorced adults is strongly and statistically signicantly correlated with all forms of suicide. A causal story from divorce to depression and suicide is plausible. Divorce also leads suicide in individual-level data (Kposowa, 2000). Depression,

10 It is worth noting that any association represents a lower bound on what the association would be if there were more constraints on acquiring new guns, such as a ban on acquisition, because then the prevailing rates would be more binding. 11 To conrm that our results are not being driven by the use of the pooled multiyear sample, Table A.2 shows results from the same models operating instead on 50-state sample which uses mean values over the three year period. Results are substantially the same, with effects of possession on suicide being slightly stronger.

12 This was calculated as the population-weighted state-average distance from each census tract (avg. population 4322) population centroid to the nearest hospital with emergency room and surgery. 13 Social connectedness is based upon answers to questions in the Civil Life in America Survey. Specically, we calculate the rst principal component of how often people answer frequently to the questions of whether they talk to or do favors for neighbors, and percentages answering yes to whether they participate in any of the following groups: School group, neighborhood or community association; service or civic association; sports or recreation association; church, synagogue, mosque or religious institution; or, any other type of organization.

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188 Table 2 BRFSS (2001/2/4) rearm ownership and suicide (baseline, minimal, and full controls). Baseline GunSui (1) guninhouse povertyRate unemployment pctUrbanLand pctUrbanPop pctM65up pctWhite gini fmd alcDrugDepAbuse medHouseIncome pctSingleMomFams pctDivorced distToHospital socialConnections lpop Obs. R2 F statistic 0.969*** (0.044) 150 0.964 134.479 0.924*** (0.031) 150 0.979 432.221 0.939*** (0.03) 150 0.984 381.533 0.958*** (0.036) 150 0.963 158.168 0.9*** (0.027) 150 0.978 391.81 0.926*** (0.027) 150 0.984 462.132 0.017*** (0.005) 0.014 (0.018) 0.05* (0.03) 0.015** (0.007) 0.001 (0.005) 0.041 (0.045) 0.819*** (0.313) 0.005 (0.022) 0.015 (0.017) 0.024 (0.026) NongunSui (2) 0.008** (0.004) 0.006 (0.015) 0.026 (0.025) 0.007*** (0.002) 0.003 (0.003) 0.089*** (0.032) 0.408** (0.193) 0.026 (0.017) 0.004 (0.009) 0.006 (0.017) Suicide (3) 0.005* (0.003) 0.005 (0.012) 0.034 (0.024) 0.008** (0.003) 0.001 (0.003) 0.066** (0.032) 0.391** (0.166) 0.014 (0.012) 0.007 (0.009) 0.012 (0.014) Minimal GunSui (4) 0.016*** (0.005) 0.012 (0.015) 0.054* (0.03) 0.015** (0.006) 0.001 (0.005) 0.033 (0.045) 0.974*** (0.313) NongunSui (5) 0.007* (0.004) 0.018 (0.013) 0.033 (0.028) 0.008*** (0.002) 0.003 (0.003) 0.082** (0.034) 0.425** (0.185) Suicide (6) 0.005* (0.003) 0.009 (0.011) 0.040 (0.024) 0.008** (0.003) 0.001 (0.003) 0.059* (0.032) 0.474*** (0.153) Full GunSui (7) 0.014*** (0.005) 0.034 (0.022) 0.018 (0.027) 0.011* (0.006) 0.004 (0.004) 0.005 (0.033) 0.906*** (0.272) 0.001 (0.016) 0.0003 (0.012) 0.005 (0.024) 0.00002 (1.00e05) 0.012 (0.008) 0.066*** (0.021) 0.049*** (0.015) 0.005 (0.015) 1.027*** (0.035) 150 0.977 297.822 NongunSui (8) 0.011*** (0.003) 0.006 (0.021) 0.003 (0.02) 0.005*** (0.002) 0.006** (0.003) 0.086*** (0.032) 0.448** (0.193) 0.019 (0.016) 0.006 (0.008) 0.002 (0.019) 5.47e06 (9.62e06) 0.018*** (0.006) 0.031** (0.013) 0.052*** (0.012) 0.004 (0.011) 0.955*** (0.028) 150 0.983 952.729 Suicide (9)

185

0.003 (0.003) 0.012 (0.014) 0.010 (0.017) 0.005* (0.003) 0.004* (0.003) 0.035* (0.02) 0.386** (0.151) 0.008 (0.009) 0.00009 (0.007) 0.002 (0.014) 1.00e05** (6.39e06) 0.006 (0.004) 0.05*** (0.012) 0.046*** (0.01) 0.002 (0.01) 0.984*** (0.021) 150 0.99 568.014

Note: OLS regressions on rearm suicide, non-rearm suicide, and overall suicide with baseline, minimal, and full specications. All regressions have year and region xed effects. * Level of statistical signicance (with S.E. clustered by state): 10%. ** Level of statistical signicance (with S.E. clustered by state): 5%. *** Level of statistical signicance (with S.E. clustered by state): 1%.

however, could also lead to both divorce and suicide so the causal structure requires further identication. A second nding of importance is that suicide increases with the distance to the nearest hospital. It seems unlikely that distance to hospital is correlated with suicide attempts (and less so when one notes that this relationship is after controlling for percent urban, the poverty rate, the unemployment rate and the other control variables). Thus, the most plausible story is that distance to hospital reduces the number of attempts that are successful. This would accord with Nicholl, West, Goodacre, and Turners (2007) nding that mortality increases with distance to hospital in the case of cardiac arrest. A cost benet analysis of hospitals should take effects of this kind into account. 4.2. Google search measure of gun ownership Let us now examine results from other measures of rearm ownership besides BRFSS, beginning with Google Searches for gunrelated terms. This measure covers the period 20042009 and we use the same model and baseline specication as before. For this and later regressions, we report results in a summary form, reporting the coefcient on the gun possession only for each regression. The coefcients on other variables are qualitatively similar. The results are shown in Table 3 after a summary of the results from the BRFSS regressions we have just discussed.

Even though the Google Gun Searches measure is completely independent and there is only one overlapping year (2004), the results are broadly similar to BRFSS, having signicant, positive associations for both gun and overall suicide. One difference is that
Table 3 Firearms and suicide summary results. GunSui OLS BRFSS Minimal Baseline Full Google Gun Searches %/Suicides By Gun Gun Poss Composite IV BRFSS Google Gun Searches Gun Poss Composite NonGunSui Suicide

0.016*** (3.29) 0.017*** (3.40) 0.014*** (3.10) 0.013*** (4.73) 0.031*** (12.76) 0.023*** (8.10) 0.020*** (3.01) 0.011** (2.45) 0.025*** (3.22)

0.007* (1.94) 0.008** (2.16) 0.011*** (3.28) 0.000 (0.07) 0.012*** (5.21) 0.007*** (2.91) 0.007 (1.51) 0.003 (0.88) 0.005 (0.91)

0.005* (1.69) 0.005* (1.78) 0.003 (1.21) 0.007*** (3.42) 0.009*** (4.40) 0.008*** (4.17) 0.006* (1.67) 0.005* (1.74) 0.010* (1.90)

Note: Regressions use the baseline set of controls unless otherwise noted, and include year and region xed effects. The IV is Field and Stream magazine circulation, except for the Google Guns Regression, which uses the Google Hunting Search IV. T statistics are in parentheses. * Level of statistical signicance (with S.E. clustered by state): 10%. ** Level of statistical signicance (with S.E. clustered by state): 5%. *** Level of statistical signicance (with S.E. clustered by state): 1%.

186

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188

4.3. Other measures over 20002009 Finally, we consider two measures of gun ownership over the complete 10 year period from 2000 to 2009. The rst of these is the percentage of suicides committed by rearm. Results can be seen under the %/Suicides By Gun row in the summary table (Table 3). Generally, the results from this are somewhat stronger than, but still in line with the BRFSS results. The overall effect of a one percentage point increase is greater, at 0.9%; despite a stronger substitution effect, the effect on the non-gun suicide rate is 1.2%, the increased effect on gun suicides is even greater, at 3.1%. The second measure of rearm ownership for the full 10-year period is the Gun Possession Composite. Although this is composed from three sources (accidental gun deaths, NICS background checks for the purchases of rearms, and the percent of suicides by rearm), we have relatively high condence in its ability to capture variation the single variable of interest, which is actual rates of rearm possession in the US populace. Results are also very similar to the BRFSS measure, although slightly higher. This can be seen under the Gun Poss. Composite row of the summary tables. A one percentage point rise in the index (equivalent to a one percentage point rise in the rate of rearm ownership) is associated with a slightly higher 0.8% rise in overall suicides. This is largely due to the strong effect on gun suicides at 2.3% (compared with 1.7% for BRFSS).14 4.4. Nonlinear effects of gun possession Like most treatments, we might expect the application of additional rearms to have diminishing effects as the prevailing rate increases. That is indeed what we nd when adding a quadratic term to the baseline regressions, we nd that it is signicant and negative. See Fig. 4, where we have shown how the marginal effects of gun possession on suicide vary with level. The existence of these nonlinearities suggests that changes in possession rates on the margin have the weakest effects. This result can explain why many studies show small or negligible effects from gun control in the United States (e.g. Ludwig & Cook, 2000) but others show more substantial effects from very large restrictions on gun ownership such as occurred, for example, in Australia (Leigh & Neill, 2010). The fact that gun control may only be effective when it results in very large decreases in gun ownership suggests that it may be difcult to square the trade-off noted in the introduction between gun ownership by law-abiding hunters and other gun enthusiasts and potentially opportunistic suicides. 4.5. Instrumental variables approach Using regression to control for other variables reveals a causal effect only if we have identied all the relevant causal variables that are correlated with gun ownership. An alternative approach to causation is to use an instrumental variable, a variable that is associated with the level of rearm ownership, but that is not related to

.02
15

Effects on linear prediction of suicides 0 .02

the Google measure does not show signs of substitution between rearm and non-rearm suicide. As a result, the point estimate of the association between guns and overall suicides is higher, at 0.7% per one percentage point increase in the index, even though the effect on Suicides By Gun is lower, at 1.3% (compared with 0.5 and 1.9 percents, respectively).

Average Marginal Effects of gunPossComposite with 95% CIs

.04

20

25

30

35 40 gunPossComp

45

50

55

60

Fig. 4. Marginal effects of guns on suicide. Note: Graph shows marginal effects of gun possession composite measure of gun ownership on suicides at varying rates of ownership using the baseline specication over the 20002009 period, with 95% condence intervals.

the suicide rate, except through the effect of guns. One such instrument might be interest in hunting. Clearly those having such an interest will be more likely to have a gun. Interest in hunting also seems to be independent of suicides, at least after accounting for the controls we are using, such as percent urban and levels of frequent mental distress. We are not aware of any evidence that interest in hunting is directly related with suicide.15 If this is correct then the exclusion restriction holds, and the instrument should not be correlated with the error term. Using interest in hunting as an instrument would then allow isolating a causal effect of guns on suicides. We use two measures of interest in hunting. The rst relies on revealed preference: circulation rates for a prominent hunting-related magazine. We gathered data on annual average circulation per 1000 inhabitants, disaggregated to the state level, for Field and Stream, a popular monthly periodical about hunting and shing (Audit Bureau of Circulations, 2012). The magazine is the most widely read magazine covering hunting in the United States by a wide margin. Baseline rates were adjusted for secular declines in overall nationwide magazine readership over the period. The second measure of hunting interest is the relative intensity of Google searches for hunting-related topics. This is available beginning in 2004 (Google, 2012), and we thus use it with the Google gun search measure of rearm possession.16 Results using two stage least squares estimation are shown in the lower section of the Summary Table 3. The rst thing to note is that in all cases, the instrument is highly signicant in the rst stage (at the 1% signicance level or above), so the instrument is indeed relevant (detailed rst stage regressions are in Table A.3). Examining the results, we see little change from OLS to IV estimates, and in no case can we reject the null of exogeneity, which makes sense given our expectations of little reverse causality for suicide. The point estimates for BRFSS and the gun possession composite are slightly higher for the 2SLS estimation compared with

14 We have investigated the variation of the impact of guns on suicide over the study period and found no strong trends; annual results are presented in Appendix Fig. A.1. We have also performed regressions using minimal and full specications for the other proxies and with IVs, and found broadly similar results to these.

15 What little evidence we can nd indicates that there is either no correlation between hunting permits and violence against humans, or a negative one (Eskridge, 1986; Flynn, 2002) 16 We do not present IV results for the %/Suicides By Firearm measure because neither of our instruments is strong in the rst stage for this measure. Also, because only one instrument is ever strong in each regression, we are unable to perform over-identication tests.

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188

187

OLS, and Google guns searches are slightly lower.17 In some cases the levels of statistical signicance are lower, reecting the lower efciency of 2SLS estimation, but qualitative results are much the same guns strongly affect the gun suicide rate, and while there are signs of substitution from gun to non-gun suicide, the overall effect remains positive. Taken together, these point estimates infer that 1% rise in the prevalence of guns causes a mean increase of between 0.5 and 1.0 percent in suicides. We take this as strong conrmatory evidence for our OLS results. 5. Conclusion In brief, we nd strong evidence that increases in gun prevalence cause an increase in rearm suicides. Despite substantial substitution of methods, we also see strong evidence that increased gun prevalence causes an increase in overall suicide. The magnitude of this result is not trivial. If all states were to reduce the rates of ownership by 10 percentage points, the expected result would be between 1640 and 2960 fewer deaths by suicide each year (59% fewer, across results from our baseline OLS specication). The mean effect obscures substantial heterogeneity. When gun ownership is very high, marginal changes in ownership levels have smaller effects. Thus our non-linear results indicate that if a state with the same high levels of gun prevalence as Alabama (at 54% ownership) but otherwise average characteristics were to reduce ownership by 10 percentage points, we would expect suicides to fall by a smaller amount than at the mean, 2.7%. Reductions in rearm prevalence are increasingly effective over the range of our data, however, so if a state with median levels of gun ownership (on par with Georgia or North Carolina at 40.7%) were to reduce ownership levels by 10 percentage points, to reach the levels currently seen in New Hampshire, suicides would be expected to fall by 13.4%. These broad results hold using a wide variety of proxies for rearm ownership, across a range of control variables, and using either OLS or IV estimation. We emphasize that we present no solution to the ecological problem. The unit of analysis is states rather than individuals which limits the quantity of data and the sources of variation we can draw upon for causal analysis. Improved measures of gun ownership and better household-level longitudinal studies would be useful to better understand the many and serious effects of rearm availability. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.irle.2013.10.004. References
Audit Bureau of Circulations, 2012]. Publishers statements. http://www. accessabc.com Azrael, D., Cook, P.J., Miller, M., 2004]. State and local prevalence of rearms ownership measurement, structure, and trends. Journal of Quantitative Criminology 20, 4362. Brent, D.A., Perper, J.A., Moritz, G., Liotus, L., Schweers, J., Balach, L., et al., 1994]. Familial risk factors for adolescent suicide: A case-control study. Acta Psychiatrica Scandinavica 89, 5258. Bureau of Alcohol, Tobacco, Firearms and Explosives, 2011]. Firearms commerce in the United States. U.S. Department of Justice, Bureau of Alcohol Tobacco and Firearms, Washington, DC http://www.atf.gov/les/publications/ rearms/121611-rearms-commerce-2011.pdf

17 Note that the hunting IV is capturing the variation in gun ownership owing to hunting. To the extent that hunters are more likely to own long guns than pistols, and the suicidal are less likely to use long guns in the absence of a pistol, then these results may underestimate the effect more generally of guns on suicide.

Bureau of Labor Statistics, 2012]. Local area unemployment statistics. http://www. bls.gov/lau/tables.htm Centers for Disease Control and Prevention, 2010]. Behavioral risk factor surveillance system survey data. U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, Atlanta http://www.cdc.gov/BRFSS/ Centers for Disease Control and Prevention, 2012]. CDC WISQARS compressed mortality & non-fatal injury les. U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Injury Prevention and Control, Atlanta http://www.cdc.gov/injury/wisqars Choi, H., Varian, H., 2012]. Predicting the present with google trends. Economic Record 88, 29. Corporation for National & Community Service, 2012]. Civic life in America. http://civic.serve.gov Della Penna, N., Huang, H., 2010]. Constructing consumer sentiment index for U.S. using google searches. University of Alberta Department of Economics Working Paper 2009-26. Demographia, 2012]. US state urban shares, population & density: 2000. www.demographia.com/db-urban2000state.htm Duggan, M., 2001]. More guns, more crime. Journal of Political Economy 109, 10861114. Eskridge, C.W., 1986]. Zero-order inverse correlations between crimes of violence and hunting licenses in the United States. Sociology and Social Research 71, 55. Federal Bureau of Investigation, 2012]. National instant criminal background check system. http://www.fbi.gov/about-us/cjis/nics Flynn, C.P., 2002]. Hunting and illegal violence against humans and other animals: Exploring the relationship. Society and Animals 10, 137154. Gallup, 2011]. Guns. http://www.gallup.com/poll/1645/guns.aspx Geneva Graduate Institute of International Studies, 2007]. Small arms survey 2007: Guns and the city. Cambridge University Press, Geneva, Switzerland. Google, 2012]. Google trends. http://www.google.com/trends Hamdi, E., Amin, Y., Mattar, T., 1991]. Clinical correlates of intent in attempted suicide. Acta Psychiatrica Scandinavica 83, 406411. Hawton, K., 2007]. Restricting access to methods of suicide. Crisis: The Journal of Crisis Intervention and Suicide Prevention 28, 49. Health Forum and ESRI, 2008]. U.S. hospitals. Vector digital data. Killias, M., van Kesteren, J., Rindlisbacher, M., 2001]. Guns, violent crime, and suicide in 21 countries. Canadian Journal of Criminology 43, 429. Kleck, G., Patterson, E.B., 1993]. The impact of gun control and gun ownership levels on violence rates. Journal of Quantitative Criminology 9, 249287. Kposowa, A.J., 2000]. Marital status and suicide in the national longitudinal mortality study. Journal of Epidemiology and Community Health 54, 254261. Krug, E., Powell, K., Dahlberg, L., 1998]. Firearm-related deaths in the United States and 35 other high- and upper-middle-income countries. International Journal of Epidemiology 27, 214221. Leigh, A., Neill, C., 2010]. Do gun buybacks save lives? evidence from panel data. American Law and Economics Review 12, 509557. Lubin, G., Werbeloff, N., Halperin, D., Shmushkevitch, M., Weiser, M., Knobler, H.Y., 2010]. Decrease in suicide rates after a change of policy reducing access to rearms in adolescents: A naturalistic epidemiological study. Suicide and LifeThreatening Behavior 40, 421424. Ludwig, J., Cook, P.J., 2000]. Homicide and suicide rates associated with implementation of the Brady Handgun Violence Prevention Act. JAMA: The Journal of the American Medical Association 284, 585591. National Academy of Sciences, National Research Council, 2004]. Firearms and violence: A critical review. The National Academies Press, Washington, DC http://www.nap.edu/openbook.php?record id=10881&page=1 Nicholl, J., West, J., Goodacre, S., Turner, J., 2007]. The relationship between distance to hospital and patient mortality in emergencies: An observational study. Emergency Medicine Journal 24, 665668. Nisbett, R.E., Cohen, D., 1996]. Culture of honor: The psychology of violence in the South. Westview Press. Owens, D., Horrocks, J., House, A., 2002]. Fatal and non-fatal repetition of self-harm a systematic review. British Journal of Psychiatry 181, 193199. Peterson, L.G., Peterson, M., OShanick, G.J., Swann, A., 1985]. Self-inicted gunshot wounds: Lethality of method versus intent. American Journal of Psychiatry 142, 228231. Rosenfeld, R., Messner, S.F., Baumer, E.P., 2001]. Social capital and homicide. Social Forces 80, 283310. Shenassa, E.D., Daskalakis, C., Buka, S.L., 2006]. Utility of indices of gun availability in the community. Journal of Epidemiology and Community Health 60, 4449. Simon, T.R., Swann, A.C., Powell, K.E., Potter, L.B., Kresnow, M.-j., OCarroll, P.W., 2002]. Characteristics of impulsive suicide attempts and attempters. Suicide and Life-Threatening Behavior 32, 4959. Skopek, M.A., Perkins, R., 1998]. Deliberate exposure to motor vehicle exhaust gas: The psychosocial prole of attempted suicide. Australasian Psychiatry 32, 830838. Sloan, J.H., Rivara, F.P., Reay, D.T., Ferris, J.A., Kellermann, A.L., 1990]. Firearm regulations and rates of suicide. New England Journal of Medicine 322, 369373, PMID: 2393410. Smith, T.W., Marsden, P., Hout, M., Kim, J., 2011]. General social surveys, 19722010 [machine-readable data le]. National Opinion Research Center, Chicago. Substance Abuse and Mental Health Services Administration, 2012]. National survey on drug use and health, national household survey on drug abuse. http://www.samhsa.gov/data/NSDUH.aspx

188

J.T. Briggs, A. Tabarrok / International Review of Law and Economics 37 (2014) 180188 Vernick, J.S., Hepburn, L.M., 2003]. State and federal gun laws: Trends for 197099. In: Evaluating gun policy: Effects on crime and violence. Brookings Institution Press, Washington, DC, pp. 345411. Vyrostek, S.B., Annest, J.L., Ryan, G.W., 2004]. Surveillance for fatal and nonfatal injuries United States, 2001. CDC Surveillance Reports. http://www.cdc.gov/ mmwr/preview/mmwrhtml/ss5307a1.htm#tab12 Wintemute, G.J., Parham, C.A., Beaumont, J.J., Wright, M., Drake, C., 1999]. Mortality among recent purchasers of handguns. New England Journal of Medicine 341, 15831589.

US Census Bureau, 2012a]. 2012 statistical abstract. http://www.census.gov/ compendia/statab US Census Bureau, 2012b]. American community surveys, census 2000, census 2010. http://factnder2.census.gov US Census Bureau, 2012c]. Census 2000 census tracts. http://www.census.gov/ geo/www/tractez.html US Census Bureau, 2012d]. Census small area income and poverty http://www.census.gov/did/www/saipe/data/statecounty/ estimates. data