We assembled an extensive panel data set from the United States including rates of suicide, antidepressant prescriptions, firearm ownership, and other covariates of interest.
Households with firearms
The proportion of adults living in households with at least one gun was obtained from the Behavioral Risk Factor Surveillance System (BRFSS) (Centers for Disease Control and Prevention (CDC) 2001), a large survey of U.S. adults conducted annually by the Centers for Disease Control and Prevention (CDC) since 1994. Designed primarily to produce state-level estimates, since 2002 BRFSS has also produced county level estimates for eligible counties. Data on gun ownership was collected from all 50 states in the years 2001, 2002, and 2004, and for 146 and 199 counties in 2002 and 2004, respectively. If an estimate for a county was available for only one year, that estimate was used. When estimates were available for both years, an averaged estimate was used. After combining the two years, 219 counties representing 49 states and the District of Columbia were available. These counties represent 7% of the 3140 counties in the United States, and 46% of the total population. Surveyors asked respondents (yes/no) if there “are any firearms kept in or around your home?” and instructed respondents to “please include weapons such as pistols, shotguns, and rifles; but not BB guns, starter pistols, or guns that cannot fire. Include those kept in a garage, outdoor storage area, or motor vehicle”. The BRFSS has not included questions on gun ownership since 2004.
Suicides
Annual counts of completed suicides for each state from 2001 through 2004, inclusive, were ascertained from the Center for Diseases Control’s National Vital Statistics System (Centers for Disease Control and Prevention (CDC)). Annual suicide counts were also collected separately by method: firearm and non-firearm. For secondary analyses, suicide counts were collected for counties from 2001 through 2005. Suicides from 2005 were included to improve the power of county level analyses. Due to suppression of small cell counts, deaths were summed over the five years. Counts were collected for total suicides and separately by method: firearm and non-firearm.
Antidepressants
Antidepressant prescription data came from the IMS Health National Prescription Audit Database. Prescription data constitute a nationally representative random sample of pharmacies (stratified by type, size and region) and capture approximately 70% of all prescriptions filled in the US. Rates were calculated as number of prescriptions per 100,000 people.
Other covariates
Additional covariates identified as potential confounders in prior work included in this study are median income (Hawton et al. 2001), unemployment rate (Lundin and Hemmingsson 2009), and the percent of the population living in an urban area (Singh and Siahpush 2002). These covariates were treated as continuous variables. Median income and unemployment rates are available annually from the Bureau of Labor Statistics. Annual estimates of the percent of population living in urban areas were calculated by assuming that the rate of change between 1990 and 2000 (available census estimate) was constant. For county-level analyses, Rural Urban Continuum Codes (RUCC) were used as an ordinal scale to measure the percent of the population living in urban areas. These codes are assigned by the US Department of Agriculture (US Department of Agriculture (USDA)).
Analysis
Mixed-effects regression models estimated incidence rate ratios and corresponding 95% confidence intervals. A negative-binomial distribution was selected to allow for over-dispersion of suicide rates about the mean. The regression models included a randomly varying intercept to allow for natural heterogeneity in suicide rates across states and to account for the correlation among the suicide rates over the five measurement occasions (2001-2004 in state-level analyses). The effect of antidepressant prescription rates (prescriptions per 100,000 people), and the effect of household firearm ownership rates (percent of households with firearms) were considered fixed effects; confounders were also treated as fixed effects. All models included dummy variables for each year in the study period to account for possible temporal trends. To distinguish cross-state and within-state variation, separate effects (of prescription rates and household firearm ownership) were estimated by including two versions of each factor in the models: (1) the mean of the factor over the four years and (2) deviations from the factor mean at each year. Specifically, the general form of the mixed models for the suicide rates was as follows:
where μij denotes the suicide rate in the ith state at the jth occasion, Fij (Aij) denotes the firearm ownership rate (antidepressant prescription rate) in the ith state at the jth occasion, and AvgFi (AvgAi) is the mean firearm ownership rate (mean antidepressant prescription rate) over the four occasions. In this model, bi is the random effect, allowing for heterogeneity in the suicide rates across states; β1 is the cross-state estimate of the effect of firearm ownership whereas β2 is the within-state estimate of the effect of firearm ownership. Similarly, β3 is the cross-state estimate of the effect of antidepressant prescriptions whereas β4 is the within-state estimate of the effect of antidepressant prescriptions. When exponentiated, the β’s can be interpreted as incidence rate ratios. Finally, we note that the fitted models also included the effects of confounders and indicator variables for time. Models were fit using PROC GLIMMIX in SAS.
For ease of interpretation of the regression coefficients, antidepressant prescription rates, household firearm ownership, and other covariates were standardized to a mean of zero and a standard deviation of one. Thus the reported incidence rate ratios correspond to a one standard deviation departure from the mean of the factor.
Secondary cross-sectional analyses were run on the available county level data aggregated over time. As with state-level analyses, negative-binomial regression models were repeated in the same pattern as primary analyses; models were fit using PROC GENMOD in SAS.
Results presented pertain to cross-state variation as suicide rates, firearm ownership, and prescription rates all changed trivially from one year to the next within any one state and no factor was significantly related to suicide rates in analyses that focused on within-state variation (not shown). Secondary cross-sectional analyses were run on the available county level data aggregated over time. As with state-level analyses, negative-binomial regression models were repeated in the same pattern as primary analyses; models were fit using PROC GENMOD in SAS.
For ease of interpretation of the regression coefficients, antidepressant prescription rates, household firearm ownership, and other covariates were standardized to a mean of zero and a standard deviation of one. Thus the reported incidence rate ratios correspond to a one standard deviation departure from the mean of the factor.