Using pooled cross-sectional time-series data for the 50 U.S. states over a 25-year period, this article examines how well four conceptual groups of social correlates—demographic, economic, social, and cultural factors—are associated with the 1976–2000 patterns in overall suicide rates and suicide by firearms and other means. Unlike past research that typically considers only one dimension, this analysis differentiates between spatial and temporal variation in suicide rates to determine whether and how social correlates operate differently in these two contexts. Results indicate that suicide rates correspond closely to social correlates. Within U.S. states, lower overall suicide rates between 1976 and 2000 were associated with demographic change (e.g., larger numbers of foreign-born) as well as with fewer numbers of Episcopalians. Across U.S. states, variation in overall suicide rates over the period was related to demographic (percentage male), economic (per capita income), social (percentage divorced), and cultural (alcohol consumption and gun ownership) factors. However, findings differ importantly by type of suicide, and across time and space. Reasons for these distinct patterns are discussed.
Suicide presents a serious public health problem in the United States as the 11th leading cause of death, accounting for more than 30,000 deaths annually (Centers for Disease Control and Prevention (CDC) 2009). This issue received a great deal of attention during the 1960s, 1970s, and 1980s as suicide rates, particularly among youth, climbed steeply. Since the mid-1980s, however, U.S. suicide rates have declined steadily, from 12.4 per 100,000 in 1985 to 10.4 per 100,000 in 2000, although these national-level figures mask geographic variation in rates of change during this time (see Fig. 1). The majority of U.S. states experienced declines between 1985 and 2000: California led in declines, with its suicide rate decreasing at an annual rate of about 3 %, from 14.3 to 8.8 per 100,000. Still, a number of states exhibited fairly stable suicide rates, and two states—Alaska and Hawaii—actually registered statistically significant increases in suicide during this time. For example, the suicide rate for Alaska rose on average by 2.9 % annually, from 14.1 per 100,000 in 1985 to 21.4 per 100,000 in 2000.
In addition to geographic variation in temporal changes in suicide, substantial geographic variation in suicide levels persists within the United States. Research has long shown that the West exhibits the highest rates of suicide compared with other regions of the country, particularly the Northeast, the region with the lowest recorded rates of suicide (CDC 2009). Figure 2 displays the average suicide rate from 1976–2000 for the 50 states. Nevada had the highest average suicide rate over this 25-year period, at 26.2 per 100,000. The corresponding figure for New Jersey, the state with the lowest average suicide rate, was just 7.1 per 100,000. Thus, U.S. suicide rates are characterized by geographic differences in both levels and change over time.
Research attempting to understand such temporal and geographic variation in suicide rates has a long tradition, beginning with Durkheim’s (1951) classic study on the subject. Although suicide risk is undoubtedly affected by individual circumstances and characteristics, such as a history of mental illness and/or substance abuse, environmental social conditions also importantly affect suicide rates (Duberstein et al. 2004a, b; Thorlindsson and Bjarnason 1998). According to Durkheim, low levels of social integration within a society lead to instability and lack of cohesion, producing excessive individualism and high rates of egoistic suicide. Lack of social regulation produces anomie, an absence of norms and an inability of society to meet the population’s needs and expectations, and corresponds to high rates of anomic suicide. As Durkheim demonstrated using European data from the 1800s and as others have shown since (for reviews, see Stack 2000a, b), the extent of social integration (usually captured by variables measuring family structure and religiosity) and social regulation (which is often proxied by economic conditions) across time and place tend to correspond in the expected fashion with suicide rates.
Such ecological studies of suicide adopt both cross-sectional and time-series research designs and thus exploit different components of overall variation in suicide rates. The connection between economic conditions, often measured with unemployment rates, and suicide rates appears stronger in time-series studies (Gruenewald et al. 1995; Luo et al. 2011; Wasserman 1984) than in cross-sectional studies. Although the majority of time-series studies support the unemployment-suicide link, cross-sectional analyses using U.S. states typically do not find support (Burr et al. 1994; Girard 1988), although the opposite is true when smaller units of analysis, such as counties, are used (Breault 1988; Faupel et al. 1987; Kowalski et al. 1987; Stack 2000a). Considerable evidence supports the notion that weak family structure, measured by the percentage divorced, is positively associated with suicide rates (Stack 2000b). However, in the case of family structure, cross-sectional analyses (Kowalski et al. 1987; Lester 1995; McCall and Land 1994) are more likely to provide support than longitudinal studies (Stack 2000b).
Research on religion is more mixed, with less division by spatial and temporal variation. Some studies have supported Durkheim’s (1951) original finding that places where Catholicism—a religion that emphasizes social integration and with strong proscriptions against suicide—is more prevalent exhibit lower suicide rates (Burr et al. 1994; Cutright and Fernquist 2004; Faupel et al. 1987). Others, however, have shown that these patterns may be attributable to misclassification of suicide deaths among Catholics (van Poppel and Day 1996) or failed to find any association after other factors are taken into account (Bankston et al. 1983; Kowalski et al. 1987). Still others (Pescosolido and Georgianna 1989) argued that different religions have network systems of varying structure and strength that account for the observed variation in suicide rates according to religious composition, but a recent study (Van Tubergen et al. 2005) failed to find support for this idea, instead observing that religious communities have a general protective effect.
Although Durkheim (1951) contended that rates of alcoholism were not linked to suicide rates, a number of aggregate-level studies, the majority of which examine variation over time, have documented a positive association between alcohol use and suicide rates (Gruenewald et al. 1995; Kalmar et al. 2008; Pridemore and Snowden 2009; Wasserman 1989). Gruenewald et al. (1995), for example, found that in the United States, a 10 % increase in alcohol consumption was associated with a 1.4 % rise in suicide rates between 1970 and 1989. Finally, some ecological studies have investigated the association between guns and suicide, in an attempt to circumvent problems associated with case–control studies (Brent et al. 1991; Conwell et al. 2002; Kellermann et al. 1992) that are limited geographically and based on small numbers of suicides (Florentine and Crane 2010; Miller et al. 2002). Miller and colleagues, using national-level cross-sectional data, found a positive association between household firearm ownership rates and rates of suicide (Miller et al. 2002; Miller et al. 2007; Miller and Hemenway 2008).
Somewhat surprisingly, there has not been, to my knowledge, a comprehensive examination of how such social conditions have influenced the declines in suicide rates in the U.S. during the latter part of the twentieth century. Several studies have explored the relationship between rising antidepressant drug use and falling suicide rates in the United States, but these investigations cover only a few years and do not include extensive controls (e.g., Gibbons et al. 2005; Grunebaum et al. 2004; Milane et al. 2006). Furthermore, as the preceding brief review attests, research on the social correlates of suicide reveals important differences in results depending on the research design used. Similar discrepancies between cross-sectional and time-series studies for other outcomes have been noted elsewhere (Beck 1980; Marvell and Moody 1991; Mouw 2002; Phillips 2006) but have not been examined systematically for suicide.
Using pooled cross-sectional time-series data for the 50 U.S. states over a 25-year period, the present study addresses these gaps in the literature and asks the following research questions. First, I examine how well four conceptual groups of social correlates—demographic, economic, social, and cultural factors—can explain the 1976–2000 temporal patterns in overall suicide rates in the United States. The declines by type of suicide have not been uniform (see Fig. 3); the firearm suicide rate actually rose somewhat, from 6.78 per 100,000 in 1976 to 7.59 per 100,000 in 1990, before declining by 22 % to the 2000 level of 5.89 per 100,000. On the other hand, the decline in the rate of suicide by means other than firearms was fairly steady over the period, from 5.57 per 100,000 in 1976 to 4.54 per 100,000 in 2000. Hence, firearm and nonfirearm suicide rates are distinguished.
Second, unlike past research that has nearly always considered only one dimension (but see Gibbons et al. 2005), I differentiate between spatial and temporal variation in suicide rates to determine whether and how social correlates operate differently in these two contexts. In other words, during this period, the factors that determine varying levels of suicide across states and the factors that explain changes in suicide rates over time within states are identified. In this way, the study aims to provide a more complete understanding of the forces shaping temporal and spatial patterns in suicide risk in the United States during the latter part of the twentieth century.
The outcome of interest is the suicide rate per 100,000, defined as the number of suicide deaths divided by the population at risk, by U.S. state and year. I obtain information on suicide victims from the National Center for Health Statistics (NCHS) (U.S. Department of Health and Human Services 1968–2000), focusing only on those who resided in the United States at the time of death. Death certificates include information on method of suicide, so a distinction can be made between suicide victims who died by firearms and those who died from other means. The denominator for the suicide death rate is the total midyear population of each state in each year, acquired from the U.S. Census Bureau (SEER 1970–2000).
During the time period of study (1976–2000), both the 9th and 10th International Classification of Diseases were used by the NCHS, but the revision does not affect the classification of suicide deaths (Anderson et al. 2001). Still, official mortality data on suicides are an underestimate of all suicide deaths, given that some suicides are misclassified as accidental or undetermined (Cooper and Milroy 1995; Pescosolido and Mendelsohn 1986). However, such errors should not alter the analysis of time trends in suicide rates; studies reveal that misclassification (e.g., coroners who underreport) in any one year tends to occur similarly in subsequent years (Cooper and Milroy 1995; Sainsbury and Jenkins 1982).
I consider a number of structural characteristics shown to be associated with variation in suicide rates, broadly classified into four groups: demographic, economic, social, and cultural factors. Suicide risk varies dramatically over the life course and across demographic groups, with the vast majority of suicide deaths in the United States occurring among white males (Maris et al. 2000). Thus, controls are included for demographic composition over time and across states, with measures of age structure (percentage aged 15 to 24 and percentage older than 65), sex composition (percentage male), and racial composition (percentage white). The population size of each state in each year, logged since its association with suicide rates is nonlinear (data not shown), is also controlled because larger places may be more socially disorganized (Faris 1955). These indicators are all readily available from the U.S. Census Bureau, which collects population data in five-year age intervals by sex and race. Finally, controls are included for the percentage of the state’s population living in urban areas and the percentage foreign-born. These measures are available in each decennial U.S. census; linear interpolation techniques, assuming constant growth over the period, were used to obtain estimates for intercensal years.
Two measures of economic conditions—unemployment rates and per capita income by state and year—are incorporated into the analysis as indicators of social regulation. State data on per capita income from 1976 to 2000 are gathered from the Regional Economic Information System (REIS) (U.S. Department of Commerce, BEA 1969–2000). The per capita income figures are converted to 1982–1984 constant dollars, using the regional Consumer Price Indices (CPI) available from the Bureau of Labor Statistics (BLS) website (U.S. Department of Commerce, BLS 1967–1999). Annual information on state unemployment levels, measured as the percentage of the civilian labor force that is unemployed, is provided by the BLS (U.S. Department of Commerce, BLS n.d.).
Several variables capturing social integration and degree of social control are included in the analyses. Weak family structure is proxied by the percentage of the state population divorced in each year. Information on state divorce levels for each census year (1970, 1980, 1990, and 2000) is available from the U.S. Census Bureau. The geographic and temporal prevalence of religious ties, a potentially important form of both social integration and regulation, is captured by four variables. The first, the religious adherence rate per 1,000, is an overall measure of religiosity. Three other measures—the percentages of the population that are Catholic, Episcopalian, or other Mainline Protestant—distinguish between key religious denominations identified by Durkheim (1951) and others (Pescosolido and Georgianna 1989; Pescosolido and Mendelsohn 1986) as offering differing degrees of social integration. Mainline Protestant faiths are distinguished from Evangelical Protestantism because studies show that a greater prevalence of the former tends to increase the suicide rate while the latter exerts a protective effect (Pescosolido and Georgianna 1989). In this study, Mainline Protestant faiths include American Baptist Churches (USA), the Evangelical Lutheran Church, the Presbyterian Church (USA), and the United Methodist Church, following the classification scheme recommended by Steensland et al. (2000). Although the Episcopal Church is considered Mainline Protestant, a separate measure is included because Pescosolido and Georgianna (1989) found that the percentage Episcopalian had a large exacerbating effect on suicide rates. These measures were downloaded from the Association of Religion Data Archives (www.thearda.com) and are available for 1971, 1980, 1990, and 2000. Linear interpolation techniques, assuming constant growth over the period, were applied to interpolate the divorce and religiosity values for intercensal years.
Finally, I include two measures related to culture: gun ownership and alcohol consumption. The General Social Survey (GSS), conducted annually since 1972 by the National Opinion Research Center (NORC) at the University of Chicago, contains a question asking whether a gun is present in the household (Davis et al. n.d.). This measure is available for each year but at the divisional level only.1 A three-year moving average of this measure is computed and used to proxy geographic and temporal patterns in gun ownership. Annual alcohol consumption at the state level is measured with a variable capturing the gallons of ethanol consumed per capita by those aged 21 and older. This measure is available from the National Institute on Alcohol Abuse and Alcoholism (NIAAA) (2011).
The dependent variable, yjt, represents the crude suicide rate (CSR) for state j in time t; denotes the intercept; and represents the estimated set of parameters for xjt, the explanatory variables for each state j and year t. The primary difference between this model and the general linear model is in the treatment of the disturbance terms. The model includes a state-specific residual, νj, which varies across states but not across time and allows for correlation among observations from the same state. The model residual, , captures random variation within states over time.
A fixed-effects model treats νj, the between-state differences, as fixed and estimable and provides estimates of only for within-state effects. In contrast, a random-effects model treats νj as independent and randomly distributed and provides estimates of that capture the combined effect of the between-state and within-state components. To distinguish the ways in which selected independent variables are associated with suicide rates over time as opposed to across states, I estimate a decomposition model. This model provides separate estimates for the effect of a covariate on the dependent variable between units (the between-unit estimator) and for the annual effects of a covariate on the dependent variable within a particular unit (the within-unit estimator) (Allison 2005; Bryk and Raudenbush 1992; Hsiao 2003; Judge et al. 1985; Phillips 2006).
The parameter measures the effect of the between-state differences, where these differences are represented by state means (all means are denoted by capital letters) for a particular characteristic over the entire period. The parameter captures the effect of within-state differences, annual state-year deviations from the overall state mean. Thus, the coefficients reveal how the covariates affect temporal variation in suicide rates (over time within states), and the coefficients indicate how factors are associated with cross-sectional variation in suicide rates (across states).
The decomposition model (Eq. (2)) incorporates a state-specific residual term, νj, which is treated as a random variable and permits correlation among observations from the same state. Because states from the same geographic region may be more similar than those located far apart, a four-category regional dummy variable is included in the models to control for time-stable characteristics of regions that may affect suicide rates. The residual error term, , allows for correlation over time among observations from the same state. The model is estimated using maximum likelihood techniques with robust standard errors.
An autoregressive error structure was applied because preliminary analyses indicated this structure to be more appropriate than one in which the correlation of observations within states is assumed to be the same (chi-square = 152.5, p < .001). I include year dummy variables in the models to account for unmeasured characteristics of time periods that affect suicide rates. Finally, exploratory analyses revealed that a few independent variables are highly correlated (r > .50). I conducted sensitivity tests by removing correlated measures most likely to be subject to multicollinearity and examining the consequences for other correlated covariates. Any notable changes in results are discussed herein. All models presented in this article contain the full set of covariates.
Table 1 displays descriptive statistics for these variables, showing the mean values as well as overall, spatial, and temporal variation in these characteristics. The mean suicide rate across the 50 states between 1976 and 2000 was 13.06 per 100,000. Suicide by firearms (8.01 per 100,000) was more common than suicide by other means (5.05 per 100,000). There is far greater variation in suicide rates across states than within states over time, as evidenced by the larger spatial standard deviation compared with the temporal standard deviation. Indeed, exploratory analyses (not shown) indicate that about 82 % of the overall variation in suicide rates is attributable to variation between states and only 18 % is attributable to temporal variation within states.
I also find considerable variation across states in demographic, economic, social, and cultural characteristics, as indicated by the size of the standard deviations. In addition, there is ample variation in these characteristics within states over time. For example, the unemployment rate varies quite substantially across states and over time, with a standard deviation of 1.21 and 1.74 percentage points across states and time, respectively.
To provide some context to these numbers, Table 2 shows the mean, minimum, and maximum values for these characteristics in 1976 and in 2000. The statistics reveal wide-ranging suicide rates across states in both periods, from states with suicide rates below 10 per 100,000 to others with rates exceeding 20. The racial and ethnic composition of states is highly variable in both years, with increasing heterogeneity (declining percentage white and increasing percentage foreign-born) over the 25-year period. Economically, the United States as a whole appears better off at the end of the period than at the beginning, although these endpoints alone cannot show the substantial fluctuation in unemployment rates during the period and the increasing mean per capita income figure masks growing income inequality. Although the mean percentage divorced doubled over the 25-year period, little overall temporal change is observed in terms of religiosity with the exception of declining percentages of Mainline Protestants. However, substantial variation in religiosity across the U.S. states over the period does exist. Finally, both alcohol consumption and gun ownership declined between 1976 and 2000, from 3.4 to 2.6 gallons consumed per capita and ownership from 53.0 % to 36.3 % of households, respectively. In sum, these general temporal trends in demographic, economic, social, and cultural factors would predict a declining suicide rate over the period.
Overall Suicide Rates
Table 3 displays the results from the decomposition model, which separates the effects of covariates on suicide rates into those between states and those within states over time. Positive coefficients suggest that the social correlate and suicide move in the same direction, while negative coefficients indicate an inverse relationship. I find a number of statistically significant differences in the magnitude and sometimes direction of association of these covariates across states as opposed to over time. Overall, these covariates explain 84.4 % and 33.3 % of the variation in overall suicide rates between and within states, respectively.2
Looking first at how these social correlates affect temporal variation in suicide, I find that shifts in demographic composition (percentage young and percentage foreign-born) are associated with changes in suicide rates. States with larger proportions of foreign-born people and smaller numbers of those aged 15–24 are more likely to exhibit lower suicide rates over time. Economic factors do not appear to affect temporal variation in overall suicide rates, but a social factor—namely, percentage Episcopalian—is positively associated with temporal variation in suicide rates. For every 1-percentage-point decline in Episcopalians, the suicide rates decline by, on average, 0.994 per 100,000 over time. Cultural factors do not affect temporal variation in the overall suicide rate, nor do changes in population size or urbanicity.
Looking at the coefficients under the Across States heading, five of the 16 factors predict cross-sectional variation and in the anticipated fashion. States with higher percentages of males exhibit higher overall suicide rates as do poorer states and those with a larger divorced population. Although changes in cultural factors, such as alcohol consumption and gun ownership, do not affect changes in suicide rates over time, they are important predictors of overall suicide levels across states. For example, a one-unit increase in gallons of alcohol consumed per capita is associated with an increase of 1.121 per 100,000 in the overall suicide rate.3
Suicide Rates by Method
Table 4 displays analogous models for the firearm and nonfirearm suicide rate. As a whole, the set of covariates explain 86.5 % and 18.4 %, respectively, of the between-state and within-state variation in firearm suicide rates. The respective figures for suicide by means other than firearms are 75.6 % and 43.8 %. The findings reveal important distinctions in the determinants of suicide rates by method, which are masked when examining only the overall suicide rate, and show the value of decomposing the total suicide rate into that committed by guns and that by other means.
Demographic characteristics are associated with both firearm and nonfirearm suicides, both across states and over time, but in markedly different ways. Firearm suicides are more likely to occur in places with larger percentages of males and a smaller young population and percentage foreign-born. However, over time within states, lower levels of the percentage young are associated with lower rates of firearm suicide rate. Firearm suicide rates also tend to be lower over time in states with higher percentages of foreign-born. In contrast, suicide by means other than firearms is more common only in states with relatively large old and young populations, and in time periods when the percentage old (p = .08) is comparatively large.
Economic factors also affect various forms of suicide somewhat differently. Spatial and temporal variations in firearm suicide are not associated with variation in unemployment levels. However, states with higher unemployment exhibit lower nonfirearm suicide rates. At the 5 % level, per capita income does not predict variation in the nonfirearm suicide rate across or within states.4 Yet, the firearm suicide rate is associated with per capita income, with a negative association across states.
States with a higher percentage of the population divorced exhibit higher suicide rates of both types, but only the nonfirearm suicide rate is positively associated with family instability over time. States with a relatively large Episcopalian population exhibit a higher firearm suicide rate. On the other hand, states with higher percentages of Catholics (p = .06) and other Mainline Protestants tend to exhibit higher suicide rates from means other than firearms. Over time, states with smaller Catholic populations tend to show lower rates of nonfirearm suicides. Lower concentrations of Episcopalians over time do not appear to affect the firearm suicide rate but are associated with reduced nonfirearm suicide rates.
Cultural factors such as alcohol consumption and gun ownership predict state variation in firearm suicide rates but not in nonfirearm suicide rates. Consistent with prior findings (Miller et al. 2007), a higher prevalence of gun ownership increases the rate of firearm suicide, as expected, but there is no concomitant negative association with nonfirearm suicide. Finally, after other state characteristics are controlled, larger states tend to exhibit higher rates of firearm suicide, but increases in population size over the period are associated with declines in firearm suicide. Variations in population size across states and over time do not appear to affect suicide by other means. After all covariates are controlled, I find no regional differences in firearm suicide rates, but the West has the highest levels of suicide by means other than firearms.
Suicide Rates by Gender and Race
Finally, given that some prior work has indicated that spatial and temporal patterns in suicide can differ by demographic group (Gunnell et al. 2003), I estimated models of overall suicide rates by gender and race, presented in Tables 5 and 6 in the appendix. Since the majority of suicide victims are white and male, the results for these two demographic subgroups largely mirror those of the overall suicide rates, both in the spatial and temporal models. Patterns are somewhat more varied for females and nonwhite suicide rates. For example, spatial variation in nonwhite suicide rates is more closely tied to demographic and economic factors than to social or cultural factors, such as percentage divorced or alcohol consumption. Only per capita income is associated (positively) with changes in nonwhite suicide rates over time.
The analyses indicate that the set of social correlates are important predictors of cross-sectional and temporal variation in U.S. suicide rates. The declines in overall suicide rates between 1976 and 2000 appear to respond to demographic shifts over the period: drops in the relative size of the young population and increases in the percentage of foreign-born are associated with decreases in suicides. In addition, lower concentrations of Episcopalians are associated with lower rates of overall suicide over time. Because both the overall suicide rate and the percentage of Episcopalians exhibit a general decline over the time period investigated (Table 2), it is reasonable to infer that states with declining numbers of Episcopalians are more likely to show drops in the overall suicide rate.
Disaggregated by type of suicide, lower rates of firearm suicide are associated with increases in the percentage foreign-born, drops in the relative size of the young population, and population growth. These covariates can explain about 18.4 % of variation over time in firearm suicides. The fall in nonfirearm suicide is correlated with declines in the percentage divorced and reductions in the percentages of Catholics and Episcopalians. Collectively, these factors explain close to one-half (43.8 %) of the temporal variation in nonfirearm suicide rates.
Demographic (percentage male), economic (per capita income), social (percentage divorced), and cultural (alcohol consumption and gun ownership) factors all predict geographic variation in overall suicide rates. However, when disaggregated by type of suicide, many more of these factors appear to explain cross-sectional variation in rates. In the case of firearm suicide rates, the majority of the social correlates is statistically significant and operates in the expected direction, explaining about 86.5 % of the between-state variation in firearm suicides. In contrast, the factors explain a bit less (75.6 %) of the geographic variation in suicide by means other than firearms. Notably, cultural factors such as alcohol consumption and gun ownership do not predict nonfirearm suicide rates across states, and some factors (e.g., unemployment rates and some religiosity variables) do not behave as expected. Taken as a whole, these factors are consistent with theoretical conjectures and work better as an explanatory model for geographic variation in firearm suicide rates and for temporal variation in nonfirearm suicide rates.
In part, these patterns may be due to displacement: the locations with more suicides by methods other than guns (wealthier states with more Catholics and Mainline Protestants (excluding Episcopalians) and more young and elderly) will be the opposite of those with high rates of firearm suicide (poorer places with more guns and alcohol consumption, a larger male population, fewer young and elderly, and relatively more Episcopalians). As a result, models of the overall suicide rate can be misleading, as revealed by analyses that decompose the overall rate by method. In this analysis, the percentage Catholic in a state is not an important predictor of either cross-sectional or temporal variation in the overall suicide rate. Yet, when the suicide rate is disaggregated by method, I find that the percentage Catholic is positively associated with both spatial (p = .06) and temporal variation in nonfirearm suicides but negatively associated with spatial and temporal variation in the firearm suicide rate (but not significantly so). These patterns are consistent with the Catholic Church’s opposition to private gun ownership and research showing that gun ownership rates vary by religious affiliation, with Protestants more likely to own handguns than Catholics and other religious groups (Little and Vogel 1992). As another example, take age composition. The overall suicide model suggests that the relative size of the older population across states and over time has virtually no effect on total suicide rates, but this is because the coefficient mixes two competing effects: the association is negative for firearm suicide rates but positive and significant for nonfirearm suicides.
The distinct effects of these correlates on firearm and nonfirearm suicide rates offer insights into the different etiology of the two types of suicide. The firearm suicide rate is closely tied to two cultural measures, alcohol consumption and gun ownership, findings that are in line with research at the individual level suggesting that firearm suicides tend to be more impulsive in nature and more closely tied to alcohol abuse (Ajdacic-Gross et al. 2006; Miller and Hemenway 2008; Simon et al. 2001). Alcohol consumption can reduce inhibitions and may increase impulsive behavior, while the ready presence of a firearm enhances the likelihood that individuals who act on impulse will complete their suicide attempt. Furthermore, the finding that gun ownership does not affect the nonfirearm suicide rate is consistent with the notion that access to means of suicide matter: substitution of an alternative method of suicide doesn’t always occur if access to guns is restricted (Miller et al. 2007). Finally, social correlates, such as the percentage divorced and per capita income or degree of urbanization (see footnote 3), are more effective in predicting changes over time in nonfirearm suicide rates. Perhaps these suicides are characterized less by factors linked to impulsive behavior but rather driven by deteriorating social circumstances, suggesting a future direction of research for individual-level studies of correlates of suicide by method.
The analysis also reveals important distinctions in effects of social correlates over time and space. In the overall suicide model, five covariates (percentage male, per capita income, percentage divorced, alcohol consumption, and gun ownership) have significantly different effects across the two dimensions. In the firearm suicide model, eight of the 16 covariates operate differently over time and space compared with three in the case of nonfirearm suicide. Some of these differences correspond to prior observations (Stack 2000a, b). For example, the cross-sectional effect of percentage divorced tends to be larger and significant in comparison to the temporal association.
Past researchers often attribute such discrepant findings to differences in units of analysis, measurement of variables, or other features of the sample, but such explanations can be ruled out here. The distinctive effects may be partly due to actual differences in how factors affect variation in behavior across place as opposed to time. The long-term (stock) effect of a variable, typically captured with cross-sectional studies, may be distinct from the transitory (flow) impact on the outcome, usually measured by time-series analyses (Kennedy 2003; Phillips 2006). The literature on poverty and well-being often makes such a distinction: many studies demonstrate that cumulative or persistent poverty is more strongly correlated with adverse health and education outcomes than short-term or recent measures of poverty (Korenman and Miller 1997). Applied to suicide, long-term exposure to adverse conditions may have a larger effect on suicide risk than short-term fluctuations, and indeed, the different effects of per capita income and percentage divorced in the overall suicide models are consistent with this idea. Similarly, alcohol consumption and gun ownership, two measures of varying cultural climates across areas, are positively associated with the overall suicide rate across states but do not predict change in overall suicide rates over time within states, perhaps because the shifts over time in culture that these variables are intended to capture are likely to be slow and incremental.
In several cases, the covariate effects on suicide rates actually operate in opposite ways across states and over time. Both the percentage male and population size tend to be positively associated with suicide rates in the cross-sectional model but negatively associated over time. States with relatively more males exhibit higher overall suicide rates presumably because males commit suicide at far greater rates than females, but it may be that a rising percentage of males within a state over time (p < .10) promotes family stability (Messner and Sampson 1991) and signifies economic prosperity—both factors that depress the suicide rate. Controls are included in the models for the percentage divorced and per capita income, but they may not adequately capture the stability and opportunity associated with an increasing male population. Although larger places tend to exhibit higher levels of suicide (e.g., between-state coefficient for population size is positive and significant in the firearm suicide model), perhaps because larger places exhibit greater levels of social disorganization (Faris 1955; Shaw and McKay 1942), growth in a state’s population over time can be viewed as a sign that the area is thriving economically and socially—forces that tend to lower the suicide rate. Similar distinct effects across time and geographic area have been observed in studies of homicide (Phillips 2006).
Nonetheless, these results must be interpreted with caution. Some of the observed differences may be ascribed simply to the fact that there is far more variation in many of the variables across states than over time. The temporal model may be less able to detect significant effects, particularly for variables such as gun ownership rates that are crudely measured (see discussion of measurement issues in the following paragraph). In some cases, the between and within estimates can be expected to be similar in the absence of omitted variables, so it is important to recognize the potential role excluded factors may play in this regard (Mairesse 1990). For example, region is sometimes a significant predictor of suicide rates across states, and if time-varying factors specific to region are excluded from the within-state model, they may contribute to some of the observed discrepancies in the between-state and within-state associations. Within states, the time period dummy variables control for omitted factors that have a uniform effect on all states over time. However, such factors may not be controlled in the between-state model of suicide rates, introducing another potential source of difference between the two estimates. Finally, it is possible that measurement error affects the between-state and within-state estimates differently. Several variables (e.g., the percentage divorced and religion variables) are interpolated for intercensus years, and this interpolation makes the differences more smoothly related to each other than they would be if based on the actual measured phenomenon. Such differences are averaged out in the between-state model but may not be such accurate measures in the within-state models.
Other forms of measurement error and omitted variable bias may also affect the findings. The measurement of gun ownership is available only at the divisional level. Unfortunately, state-level measures of gun ownership (such as the Behavioral Risk Factor Surveillance Survey (BRFSS)) are available only for a nonrandom sample of states for select years, and the Cook Index (number of gun homicides and gun suicides as a proportion of total deaths from lethal violence; Cook 1979) is a problematic measure in this analysis given that the firearm suicide rate is used as a dependent variable. However, exploratory work that substituted the state-level Cook Index for division-level household gun ownership reached similar substantive conclusions (results available upon request), suggesting that the bias is not substantial.
A notable omitted variable that has undoubtedly influenced the risk of suicide over the period is antidepressant drug usage, which rose dramatically during the latter portion of the time period studied and presumably led to greater treatment of major depression, which in turn may have lowered suicide rates. This has been the primary focus of recent time-series studies (Gibbons et al. 2005; Grunebaum et al. 2004; Milane et al. 2006), but state-level measures of this variable are not available for the time period. Measures of norms and/or attitudes toward suicide across states, shown to be a predictor of suicide (Cutright and Fernquist 2004), and other indicators of economic strain, such as state foreclosure rates (Stack and Wasserman 2010), are unavailable as well. Future work might also consider whether the presence and strength of statewide suicide prevention programs are related to trends in suicide. Finally, cautions must be raised regarding the ecological fallacy—aggregate-level relationships may not reflect individual behavior. Although the analysis shows that gun ownership and percentage Episcopalian are both positively associated with firearm suicide, one cannot say whether it is those individuals in households with guns and who are Episcopalian who commit suicide.
These limitations notwithstanding, the present analysis and methodological approach offer new insights into aggregate patterns of U.S. suicide. First, there are differences in the cross-sectional and longitudinal effects of important covariates on suicide rates, distinctions that are missed by estimates from either random-effects or fixed-effects models. Second, the approach mitigates somewhat the potential problem of omitted variable bias and thus better isolates the true effect of each social correlate on the suicide rate. Although I cannot control for all possible factors, the within-state estimates do control for time-stable state characteristics and the inclusion of year dummy variables control for omitted variables that have a uniform effect on all states over time. Some of the influence of antidepressant drug use changes over the period, for example, may be picked up by these controls. Finally, the time-series portion of the analysis yields more reliable estimates than those from national-level time-series work, given that it uses 50 replications (corresponding to the 50 U.S. states) of the time series.
A clear conclusion from the analysis is that both geographic and temporal variation in suicide rates are closely tied to a number of varying social conditions across states and over time, just as Durkheim (1951) demonstrated in nineteenth-century Europe. These findings raise cautions about work that restricts the number of covariates considered, such as the fairly exclusive focus in the psychiatry literature on rising antidepressant drug use as a cause of declining suicide rates (e.g., Milane et al. 2006) or the recent study linking business cycles to suicide rates in the United States without controls for other changing social or cultural conditions (Luo et al. 2011). Rather, a confluence of factors appears to shape suicide trends, indicating that a multipronged population-based approach to suicide prevention is warranted. The results presented here suggest that programs that restrict access to means such as firearms, promote among primary care physicians the importance of screening for major depression, and provide social support and counseling services during periods of economic recession may all be effective in reducing suicide rates. Indeed, studies that investigate such prevention strategies find support for their efficacy (see Mann et al. (2005) for a review of this research).
Moreover, the study findings highlight the ways in which firearm and nonfirearm suicides are affected differently by both demographic and social context. Moving forward, researchers must pay close attention to how the rapidly changing demographic and social context of the United States will affect overall suicide patterns, as well as the relative distribution of suicide by firearms and other means.
I thank David Greenberg, Ellen Idler, Lauren Krivo, D. Randall Smith, the anonymous reviewers, and the Editor for very helpful comments. This research was supported in part by the American Foundation for the Prevention of Suicide.
The U.S. census groups states and the District of Columbia into nine census divisions. For more detail, see the Census Bureau website (http://www.census.gov/geo/www/2010census/gtc/gtc_census_divreg.html).
The percentage of variation between and within states explained by the covariates is determined by estimating a baseline model (regional and year dummy variables only) and comparing the estimate of random variation between states and within states to that obtained when the full set of covariates is included. For example, the estimate of state-level variance in the baseline model is 4.3392 but is reduced to 0.7133 (84.4 %) when the covariates are added to the model.
Although not statistically significant in the full model, the religious adherence rate is negatively and significantly associated with overall suicide rates across states when the percentage divorced (coefficient = −0.008; SE = 0.003) or alcohol consumption (coefficient = −0.007; SE = 0.002) is removed.
Average per capita income and percentage urban are highly correlated. In the between-state model of nonfirearm suicides, per capita income is statistically significant (coefficient = 0.286; SE = 0.13) when percentage urban is removed. Percentage urban is statistically significant (coefficient = 0.037; SE = 0.02) when per capita income is removed.