This spatial panel analysis used data for 3,131 US counties from 50 US states over 17 years (2000 to 2016), for an overall sample of 53,227 county-year units.

The outcome of interest was counts of suicides. We accessed Compressed Mortality Files from the Centers for Disease Control and Prevention for county-years from 2000 to 2016. Suicides were deaths classified as International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes X60-X84 (intentional self-harm), U03 (terrorism involving intentional self-harm), and Y87.0 (late effects of intentional self-harm).

The main exposures were casinos and other gambling outlets. Other gambling outlets comprised establishments primarily engaged in operating gambling facilities or providing gambling services that are not casinos or casino hotels. Examples include but are not limited to bingo halls, off-track betting shops, slot machine parlors, and card rooms.²⁶ Using County Business Patterns data from the US Census Bureau, which documents the number of businesses present per county-year according to North American Industry Classification System (NAICS) codes, we calculated counts of casinos (NAICS codes 721120, 72112/, 731210, 73121/, and 71322//) and other gambling outlets (codes 713290 and 7132//). Consistent with previous analyses,²⁷ we denominated these count variables by county-year populations to generate measures of casino density, expressed in terms of venues per 100,000 population. These measures were heavily positively skewed (e.g., casinos: mean = 0.6, SD = 7.2, skewness = 39.0; kurtosis = 1191.4). Due to concerns that county-years with large numbers of casinos (e.g. Clark County, Nevada) would inordinately affect the estimated associations and many other county-years had zero casinos or other gambling outlets, we square root transformed both measures.

Many county-year characteristics may be associated with the presence of casinos and other gambling outlets, causally related to suicide incidence, and do not lie on the causal path between the two.^28–30 These characteristics may therefore confound relationships between the main exposures and the outcome.³¹ County Characteristics Intercensal Population Estimates from the US Census Bureau provided estimates of population composition within categories of sex, age, and race/ethnicity. We calculated proportions that were male, aged 0 to 24, aged 25 to 44, aged 45 to 64, aged 65+, non-Hispanic Black, non-Hispanic White, Hispanic, and Native American. Similarly, the US Census Bureau’s Small Area Income and Poverty Estimates program estimated the annual median household income. From the Bureau of Labor Statistics, Local Area Unemployment Statistics described the proportion of the population aged > 16 years who were unemployed, and the Quarterly Census of Employment and Wages described the local workforce according to employment within industry categories. We calculated the proportion of the workforce employed in agriculture and mining, leisure and hospitality, financial services, and other industries. Finally, we calculated time-invariant measures of land area (square miles, which can be interpreted as a measure of population density because the model offset was population) and 2010 classifications of Rural-Urban Continuum Codes from the US Department of Agriculture, separated into metropolitan (RUCC: 1 to 3), urban (RUCC: 4 to 6) and rural (RUCC: 7 to 9).

Bayesian conditional autoregressive Poisson models related casinos and other gambling outlets to counts of suicides within the 53,227 space-time units. Models were specified as: Yi,t∣μi,t~Poisson(Ei,texp(μi,t)) where Y_i,t is the observed count of suicides within county i in year t, and E_i,t is the expected number of suicides based on county-year populations. The parameter μ_i,t therefore estimates the incidence rate ratio of suicides within county i in year t. We modelled the log of the incidence rate ratio linearly: μi,t=β00+β01+λt+X`i,tβ1+ωt+θi+φi where β<sub>00</sub> is the overall intercept term, β<sub>01</sub> is the random intercept for state, and λ is a global linear time trend estimating the annual proportional change in suicide incidence over the 17 years. X` is a matrix of independent variables, including the measures of casino and other gambling venue density, and the included covariates. β₁ is a vector of fixed-effects estimating the relationships between these variables and suicide incidence. The random effect ω captures residual variance in suicide incidence across years. The parameter θ_i is a time-invariant random effect that estimates the spatially unstructured error for counties and accounts for over-dispersion of the dependent variable. A time-invariant conditional autoregressive (CAR) random effect, φ_i, captures the spatially structured error. In doing so, it controls for the loss of unit independence due to spatial autocorrelation and addresses the small area problem by borrowing strength from adjacent counties. The CAR random effects were identified using a binary matrix of adjacent counties with queen’s contiguity. Models were estimated using R-INLA, which calculates the Integrated Network Laplace Approximation of a fully Bayesian model.^34–36

In addition to the main analysis, in which the exposures of interest were the square root of casinos and other gambling outlets per 100,000 population, we conducted several sensitivity analyses to limit the possibility that our results were artifacts of our model specification or variable construction. Specifically, we conducted 5 sensitivity analyses. The first (S1) included an interaction term between casinos per 100,000 population (square root) and other gambling venues per 100,00 population (square root). The second (S2) replaced the main exposures with casinos and other gambling outlets per 100,000 population (i.e., not square root transformed). The third (S3) replaced the main exposures with unadjusted counts of casinos and other gambling outlets. The fourth (S4) replaced the main exposures with dummy variables indicating the presence of 0, 1, 2 or 3, or 4+ casinos and other gambling outlets. Lastly, the final sensitivity analysis (S5) removed Nevada counties.

Table 2 shows the results of the main Bayesian conditional autoregressive Poisson model relating the square root of casinos and other gambling outlets per 100,000 population to suicides. After controlling for confounding, the incidence rate ratio for casinos was 1.016, and the credible interval was between 1.010 and 1.023, which means that each square root increase in the number of casinos per 100,000 population was associated with a 1.6% increase in suicides per population and that the range for this estimate was between 1.0% and 2.3% increase (95% CrI: 1.010, 1.023). Other gambling outlets were positively associated with suicides after controlling for confounding, with an estimated association slightly weaker than that for casinos and including the null value of 1.000 (IRR = 1.003; 95% CrI: 0.997, 1.009).

Results of the sensitivity analyses were generally consistent with the results of the main model (see Supplemental Table 1). Model S1, which included an interaction term, found no significant interaction between the square root of casinos per 100,000 population and the square root of other gambling venues per 100,000 population. In Model S2, the number of casinos per 100,00 population was associated with a 1% increase in suicides per population (95% CrI: 1.000, 1.002) and the number of other gambling outlets was associated with a 3% increase in suicides per population (95% CrI: 1.001, 1.004) after controlling for confounding. Not adjusting for population in Model S3 yielded null results. When including the exposure as a categorical variable in Model S4, counties with four or more casinos had the largest increase in suicides per population compared to counties with no casinos, after controlling for confounding (IRR = 1.048, CrI: 1.028, 1.068). Other gambling outlets yielded null results. Finally, excluding Nevada in Model S5 demonstrated that each square root increase in the number of casinos per 100,000 population was associated with a 1.4% increase in suicides per population (95% CrI: 1.008, 1.021). The results for other gambling outlets were not statistically significant.

This study has several strengths which distinguish it from the limited literature on this topic to date. It is the first study to estimate the relationship between casinos and other gambling venue density and suicide mortality on the national scale. Previous studies have compared carefully selected casino locations with control locations at a variety of spatial scales, employing methods like cluster analysis or matching to match cases to controls.^24,25,37,42 Second, it is the first longitudinal study of the relationship between casino density and suicide. Our spatial panel design compared associations both within and between counties over time and controlled for spatio-temporal autocorrelation, thereby providing more scientifically robust estimates than a cross-sectional ecological analysis. Future studies may consider a time series analysis, which could help with confounding but would only consider within-county comparisons. Finally, this study undertook sensitivity analyses to ensure that its findings were not the result of outliers like Clark County and other Nevada counties biasing national results, or that the findings were not reliant on idiosyncratic model specifications.

Nevertheless, the study is subject to various limitations. Data were not available for some important variables, some of which may confound the relationship of interest. For example, the presence of online gambling, knowledge of an individual’s co-morbidities (i.e., substance abuse^43,44), or alcohol availability at a casino or other gambling outlet all may increase or decrease the risk of suicide. It is possible that these unmeasured variables resulted in an overestimation of our estimates. However, we included controls for key socio-economic and geographic variables which are the most substantial known drivers of geographic variation in suicide. As such, any bias is likely minimal. In addition, we were limited to counties for our unit of analysis as the County Business Patterns database documents the number of businesses present per county-year. Counties represent artificial administrative boundaries, and it is likely that the impact of new casinos and other gambling outlets will span multiple geo-administrative areas. Further, the county-level casino counts exclude race-course casinos, or ‘racinos’. According to industry statistics, racinos accounted for 5.3% of all casinos in the US in 2017. The proportion was higher in specific states like Ohio where seven of the state’s eleven casinos are at race tracks.⁵ The external validity of the County Business Patterns database for other types of casinos has yet to be studied. Next, suicide data are subject to under-reporting. If under-reporting rates are associated with casino density, then our results could be biased towards the null. Finally, we made no attempt to model potential spatial or temporal non-stationarity of the relationship between casino density and suicide mortality. Given the variety of regulatory structures in the US, it is likely that this relationship will vary to some degree from place to place and indeed may vary over time as predicted by adaptation theory. Yet, due to an inability to sufficiently control for this non-stationarity, we have estimated the average relationship across all county years.

Casino and other gambling venue densities are both associated with suicide mortality in the US between 2000 and 2016. While quasi-experimental and other more rigorous study designs will better support causal inference, this preliminary evidence is consistent with clear theoretical mechanisms linking casinos and suicides. Potential impacts on suicide should be considered when legalizing casinos and other gambling venues.