Socio-Spatial Patterns of Suicide Mortality in the United States

Abstract

Suicide mortality in the United States exhibits substantial geographical and sociodemographic heterogeneity. Yet the role of large-scale social networks in shaping this variation remains underexplored. We integrate data on county-level suicide mortality (2010–2022) and Facebook’s Social Connectedness Index (SCI) to assess how both the risk of suicide mortality and the effect of firearm restriction policies propagate through inter-county social ties. First, using two-way fixed effects regression models with sociodemographic, economic, and spatial controls, we find that a one-standard-deviation increase in (SCI-weighted) suicide mortality in socially connected counties is associated with an increase of 2.78 suicide deaths per 100,000 people in the focal county (95% CI: 1.06– 4.50). Second, we examine Extreme Risk Protection Orders (ERPOs) — state-level firearm policies that allow temporary restriction of firearm access for individuals at risk of self-harm — and show that counties with stronger (Facebook) social ties to ERPO-adopting states experience reductions in suicide mortality, even without local policy implementation. Our findings suggest that a one-standard-deviation increase in ERPO social exposure is associated with a decrease of 0.301 suicide deaths per 100,000 people in the focal county (95% CI: 0.480–0.121). This protective association persists after adjusting for geographical proximity and including state-by-year fixed effects that capture time-varying state-level factors. In sum, our findings suggest that social networks can facilitate the diffusion of both harmful exposures and protective interventions. This socio-spatial structuring of suicide mortality underscores the need for network-driven prevention strategies that incorporate social network topology (e.g., SCI-derived influence metrics), alongside more traditional approaches based on geographical targeting.

INTRODUCTION

Suicide mortality is a persistent public health crisis in the United States (US). In 2023, suicide was the eleventh leading cause of death in the US, accounting for an estimated 49,000 fatalities [1]. However, this national statistic does not capture substantial age-specific heterogeneity. Adolescents and working-age adults are the most vulnerable. Among individuals aged 10–34 and 35–44, suicide was the second and fourth leading cause of death, respectively [1]. To better understand and address these disparities across different demographics, the social ecological framework provides a valuable lens, categorizing risk factors and preventive measures for suicide at the individual, community, or societal level [2]. Examples of risk factors include social isolation [3], mental health and substance use disorders [4, 5], bullying and cyberbullying [6], and access to lethal means (particularly firearms) [7]. Further, knowing someone who died by suicide [8], or being exposed to suicide-related information through social media may amplify one’s suicide risk [9].

In this paper, we focus on the effect of social networks, which span the individual level (close interpersonal connections) and the community level (indirect and broader social ties). Social networks can either amplify or mitigate the risk of suicide, depending on their structure and nature. For instance, social isolation or interaction with negative peers may influence an individual toward unhealthy coping mechanisms, such as substance use and misuse. In addition, exposure to suicide death can increase suicidal ideation and increase capability via exposure to method-specific examples of lethal means that guide planning [10] – a phenomenon termed suicide contagion [11, 12, 13]. Durkheim [14] posited that suicide mortality rates in a given population are affected by the degree of social integration, with elevated levels under conditions of social disconnection. Building on this foundational theory, empirical studies have demonstrated that the risk of suicide mortality increases following exposure to suicidal behavior within one’s social network [15]. In sum, theoretical and empirical studies suggests that suicidal behaviors can propagate through both interpersonal and community networks [16, 17, 13]. The role of such social exposures is especially prominent among adolescents and young adults who frequently encounter suicide-related content via digital media [18, 19]. Despite these negative influences, social connection is protective: family, peer, and school connectedness and a sense of belonging are associated with lower suicidal ideation and attempts, and simple connection-building interventions (e.g., caring contacts) reduce subsequent suicide risk [20, 21, 22, 23, 24, 25]. Given the established role of social ties in suicide risk, adopting a network science perspective to examine suicide mortality offers a valuable interdisciplinary approach for understanding complex public health issues such as suicide [26].

Until recently, efforts to assess population-level associations between social networks and suicide mortality were limited by data constraints. Researchers relied on small samples or qualitative studies due to the lack of high-resolution social network data. The release of the Social Connectedness Index (SCI) by Meta in 2018 enabled the quantification of social ties between geographic regions using aggregated Facebook friendship data [27]. The SCI measures the relative probability that a Facebook user in region i is a friend of a user in region j, normalized by the product of the number of active Facebook users in each region. This formulation allows for comparisons of social connectedness between pairs of regions, where higher SCI values indicate stronger interpersonal ties. Constructed from billions of anonymized friendship links and periodically updated, the SCI provides a scalable proxy for real-world social connectedness, facilitating the study of how network structure relates to spatial patterns in behavioral and health outcomes. Bailey et al. [27] demonstrated the empirical relevance of SCI by showing that it correlates with inter-county migration patterns, trade volumes, job search behavior, patent citations, and specific public health outcomes. The availability of SCI data has made possible studies examining how behavioral and health outcomes propagate across socially connected regions. For instance, Charoenwong et al. [28] found that counties socially connected to early COVID-19 outbreak areas exhibited earlier reductions in mobility. Holtz et al. [29] showed that compliance with public health mandates varied systematically with SCI-based social proximity to counties with more stringent policies. Similarly, Tiwari et al. [30] demonstrated that opioid overdose death rates in socially proximate counties were positively associated with rates in the focal county. Together, these findings underscore the utility of SCI to capture the diffusion of behavioral and outcomes across social networks.

Further, research shows that social network structure is a modifiable target for suicide prevention—specifically, that increasing supportive ties to low-risk peers, disrupting high-risk clustering, and leveraging prosocial central connectors (e.g.,gatekeeper programs) can shift norms and facilitate help-seeking, thereby reducing suicidal ideation, attempts, and deaths [31, 32]. However, most studies have been restricted to small samples or specific settings (e.g., school-based cohorts, military units, clinically high-risk youth) and have not evaluated whether large-scale inter-county friendship networks across the United States are associated with geographic variation in suicide mortality or with the diffusion of preventive policies. As a result, the broader socio-spatial dynamics of suicide risk–particularly the indirect transmission of protective interventions through social ties–remain insufficiently understood. One prominent example of a socially mediated preventive intervention is the implementation of Extreme Risk Protection Orders (ERPOs), also known as “red flag laws”. ERPOs are civil court orders that allow temporary restriction of firearm access from individuals deemed to present a risk to themselves or others [33]. This socially mediated approach leverages interpersonal networks to identify and minimize suicide risks proactively [34]. Specifically, it enabling law enforcement, family members, friends, and close associates to file petitions based on observed behaviors that indicate a heightened risk of suicide, which may result in temporarily restricting firearm access. Given that concerned individuals often initiate ERPO interventions within a person’s social network, they offer one potential avenue to examine how social ties influence suicide prevention efforts empirically. Empirical evidence from multiple states shows ERPOs’ efficacy in reducing suicide mortality. Connecticut’s issuance of 762 ERPOs between 1999 and 2013 was estimated to avert per 10 to 20 orders, with individuals targeted by these orders initially exhibiting suicide rates approximately 40 times higher than the general population [34]. These features make ERPOs a useful case to examine in our study, where we assess whether exposure to such policies through social connectedness influences suicide mortality across counties. Building on these insights, our study systematically evaluates the role of social connectedness in shaping county-level suicide mortality.

The contributions of our study, which integrates data on county-level suicide mortality and social connectedness, are three-fold:

1. We define and quantify county-level social exposure and spatial exposure to suicide mortality.

2. We estimate associations between county-level suicide mortality and social exposure, with and without adjustment for spatial exposure.

3. We evaluate whether network-mediated exposure to a firearm-restriction policy is associated with reduced suicide mortality in non-adopting counties that are socially connected to counties that adopted the policy.

We test two hypotheses to evaluate whether network-mediated exposures are associated with county-level suicide mortality in the US:

• H1 a one-standard-deviation increase in the SCI-weighted average suicide death rate in socially connected counties is positively associated with the focal county’s suicide death rate, controlling for spatial exposure.

• H2 a one-standard-deviation increase in social exposure to firearm-restriction policy firearm-restriction policy is negatively associated with the focal county’s suicide death rate.

We estimate county–month panel models with county and state–year fixed effects, controlling for time-varying sociodemographic covariates. Although the associations identified in our study do not imply causality, they highlight how social networks can function as pathways for both detrimental socio-spatial influence and beneficial dissemination of preventive policies. In conclusion, our findings highlight the role of social networks in suicide mortality and support the integration of network-based approaches into the design of public health interventions.

RESULTS

Socio-spatial patterns of suicide mortality in the US (H1)

We analyzed county-level suicide death data from the National Vital Statistics System (NVSS) Multiple Cause of Death files for the period 2010–2022, encompassing 40,794 county-year observations across the US. Our primary outcome of interest was the overall suicide mortality rate per county (per 100,000 people), across all age groups. Our primary objective was to assess whether exposure to suicide deaths in socially connected counties is positively associated with suicide mortality in the focal county (H1). To test whether the estimated association between social connectedness to other counties and suicide mortality in the focal county was confounded by geographical distance, we compared two approaches, with and without controlling for spatial proximity. To that end, we considered a spatial exposure variable defined as the weighted average of suicide mortality rates in all US counties other than the focal county, with county-level weights inversely proportional to the geographical distance to the focal county’s centroid.

To estimate these associations, we utilized two-way fixed effects regression models with county and year indicators, effectively controlling for time-invariant local characteristics and for annual patterns common across all counties. Exposure to suicide mortality outcomes in other counties was captured through two metrics. The first metric, denoted by s_−it, corresponds to “deaths in social proximity”. It is defined as the weighted average of suicide mortality rates in year t in counties other than focal county i, where the weight of a given county is equal to its SCI with the focal county. The second metric, denoted by d_−it, corresponds to “deaths in spatial proximity”. It is defined as the weighted average of suicide mortality rates in year t in counties other than focal county i, where the weight of a given county is equal to the inverse of its geographical distance to the focal county’s centroid. Detailed mathematical formulae for these metrics are provided in equations 1 and 2 in the Methods Section.

First, we estimated the relationship between suicide mortality in the focal county and “deaths in social proximity” without controlling for spatial proximity. In this model, a one-standard-deviation (1-SD) increase in “deaths in social proximity” was associated with an increase of 3.34 suicides per 100,000 people in the focal county (cluster-robust 95% CI: [1.76, 4.93], p < 0.01; Fig. 1, red estimate). Given that the strength of social connections is correlated with geographical proximity, we next evaluated whether this observed population-level association was confounded by spatial proximity.

Figure 1:

Role of social ties in county-level suicide mortality. Estimated regression coefficients (${\widehat{\zeta }}_1$) for suicide death rates in socially proximal counties (s_−it) in two models. Model 1 (red): without adjustment for deaths in spatial proximity (d_−it. Model 2 (blue): with adjustment for deaths in spatial proximity (d_−it). Horizontal lines denote 95% confidence intervals. The vertical dashed line indicates the null hypothesis (ζ₁ = 0). Point estimate in model 1: 3.34 (cluster-robust 95% CI: [1.76, 4.93]); point estimate in model 2: 2.78 (cluster-robust 95% CI: [1.06, 4.50]). Both models include county and year fixed effects and sociodemographic control variables (see Table 1).

Second, we estimated a model including both variables, i.e., “deaths in social proximity” and “deaths in spatial proximity”, to disentangle their independent associations with suicide mortality. In this combined model accounting for both social and spatial proximity, the association between the suicide mortality rate of the focal county and “deaths in social proximity” remained statistically significant and substantial (2.78 suicides per 100,000, cluster-robust 95% CI: [1.06, 4.50], p < 0.01; Fig. 1). Simultaneously, the suicide mortality rate of the focal county was also positively associated with “deaths in spatial proximity”, albeit at a smaller magnitude (0.88 suicides per 100,000, SE = 0.316, p < 0.05). This finding suggests that social ties independently influence the risk of suicide mortality, beyond spatial proximity.

Complete regression results, including the covariate effects of population density, age structure, racial/ethnic composition, median household income, unemployment, educational attainment, and English proficiency, are detailed in Table 1. All continuous variables in the model, including population density, median household income, and the two exposure metrics, were standardized to have mean zero and unit variance to facilitate the interpretation of estimated coefficients. Consistent with national surveillance and county-level analyses, population density and median household income are inversely associated with suicide mortality, and larger Asian and Hispanic population shares are protective; their statistical significance in Model 2 (all p < 0.05) strengthens confidence in our model specification [35, 36, 37].

Table 1:

Estimates of socio-spatial influences on county-level suicide mortality obtained via two-way fixed effect regression. Column (1) presents estimates from a model regressing county-level suicide mortality on standardized deaths in social proximity (s_−it). Column (2) additionally controls for standardized deaths in spatial proximity (d_−it) to disentangle social influence from geographical proximity. Both models include county and year fixed effects and adjust for population density, age distribution (percent aged 0–17, 18–44 and 45–64), racial composition (percent Asian, Black, and Other racial subgroups), ethnicity (percent Hispanic), median household income, percent with limited English proficiency, percent unemployed, and percent with less than high school education. Standard errors are clustered at the state level.

	Dependent variable: County-level suicide deaths per 100k
	Model 1	Model 2
Deaths in social proximity s—it	3.343*** (0.789)	2.782*** (0.858)
Deaths in spatial proximity d—it		0.777** (0.319)
Population density	−1.181*** (0.374)	−0.989** (0.376)
Percent aged below 18	−0.191 (0.347)	−0.183 (0.340)
Percent aged 18–44	0.325 (0.544)	0.131 (0.536)
Percent aged 45–64	−0.816*** (0.296)	−0.851*** (0.285)
Percent Asian	−0.838*** (0.293)	−0.822*** (0.290)
Percent Black	−1.372* (0.743)	−1.407* (0.763)
Percent Other	0.370** (0.181)	0.308* (0.163)
Percent Hispanic	−3.704*** (0.795)	−3.586*** (0.803)
Median household income	−0.698*** (0.166)	−0.692*** (0.169)
Percent with limited English proficiency	−0.067 (0.075)	−0.036 (0.072)
Percent unemployed	0.017 (0.158)	0.028 (0.150)
Percent with less than high school education	−0.013 (0.124)	−0.023 (0.123)
Observations	40,794	40,794
R2	0.946	0.946
Adjusted R2	0.941	0.941

Robust standard errors in parentheses.

^*p<0.1;

^**p<0.05;

^***p<0.01

To test the robustness of the estimated associations to outcome specification, we fitted the models again using age-adjusted suicide mortality rates as the dependent variable, thereby accounting for cross-county differences in population age structure. Results from these sensitivity analyses (see Supplementary Results; Table S1) confirmed the positive association between suicide mortality in the focal county and deaths in social proximity, although the estimated magnitudes were about 60% smaller than when using unadjusted suicide mortality rates as the outcome. Such an attenuation was expected, as age adjustment removes demographic variability in the baseline risk of suicide of a county’s population. Yet the consistency in the directionality of the estimated associations across outcome specifications underscores the robustness of our findings.

The subsequent subsection explores whether exposure to firearm restriction policies similarly propagates through socio-spatial channels (H2).

ERPO policy exposure through social networks (H2)

In this section, we investigate whether suicide mortality is associated with indirect exposure to firearm access restrictions through social networks connecting a county of interest to counties in other states that have enacted ERPOs. We hypothesize that counties with greater social connectedness to ERPO-adopting counties experience lower suicide mortality, even in the absence of local ERPO adoption. To test hypothesis H2, we construct a standardized metric of ERPO social exposure using the SCI, which quantifies cross-county social ties. We then estimate the association between ERPO social exposure and suicide mortality using a two-way fixed effects regression, accounting for county- and time-specific heterogeneity. Finally, to disentangle the role of social exposure from that of geographical proximity (and assess the robustness of our results), we incorporate two regressions, i.e., with and without control for the same spatial exposure variable to assess robustness. The analysis proceeds by first describing ERPOs and their empirical effectiveness, then estimating their direct effects, followed by results for indirect social exposure and robustness to spatial confounding. The staggered implementation of ERPOs across states provides a chance for a statistical framework to explore the effect of social networks on suicide rates.

Since ERPO implementation is specific to each state, its direct impact is limited to counties within the states that enact it. This geographic separation of direct effects allows for a thorough evaluation of indirect social network influences. Specifically, the absence of direct ERPO exposure in non-implementing states enables us to isolate and measure the indirect effects of social networks on suicide mortality, which arise solely from social connectivity rather than direct legislative actions.

We begin by estimating the association of the ERPO policy in the counties of the implementation states with suicide rates. Table 2 presents the estimated coefficients from the direct-effect specification as shown in equation 4 in the Methods section. The ERPO indicator exhibits a statistically significant association of $\widehat{\psi }$ = −0.53 (cluster-robust 95% CI: [−0.93, −0.13], p < 0.01). Standard errors are clustered by state (51 clusters), providing conservative inference that is robust to within-state correlation. Since the ERPO variable changes only at the state–year level, we include county fixed effects to control for county-specific characteristics that do not change over time, and year indicators to account for nationwide shocks such as macroeconomic cycles.

Table 2:

Estimated effects of ERPO policy exposure on county-level suicide mortality (deaths per 100,000 people). Column (1) reports direct effects of local ERPO adoption with county and year fixed effects. Column (2) reports indirect effects of ERPO social exposure, measured through cross-county social ties, estimated with county and state–year fixed effects. Column (3) reports the indirect social exposure model with an additional control for ERPO spatial exposure, also estimated with county and state–year fixed effects, to assess robustness to spatial confounding. All models adjust for population density, age distribution (percent aged 0–17, 18–44 and 45–64), racial composition (percent Asian, Black, and Other racial subgroups), ethnicity (percent Hispanic), median household income, percent with limited English proficiency, percent unemployed, and percent with less than high school education. Standard errors are clustered at the state level.

	Dependent variable: Deaths per 100K
	Direct Effect (1)	Indirect Social Network Exposure (2)	Indirect Social Network Exposure (controls for Spatial Exposure) (3)
ERPO	−0.526** (0.199)
ERPO Social Exposure		−0.214*** (0.064)	−0.301*** (0.089)
ERPO Spatial Exposure			0.525 (0.315)
Population density	−1.442*** (0.381)	−0.637 (0.603)	−0.674 (0.582)
Percent aged below 18	−0.108 (0.365)	−0.213 (0.378)	−0.196 (0.380)
Percent aged 18–44	0.448 (0.557)	−0.223 (0.550)	−0.139 (0.546)
Percent aged 45–64	−0.951*** (0.329)	−0.581 (0.388)	−0.518 (0.401)
Percent Asian	−1.000*** (0.343)	−0.668** (0.292)	−0.644** (0.275)
Percent Black	−1.298* (0.735)	−2.906*** (0.682)	−2.813*** (0.675)
Percent Other	0.477** (0.199)	−0.106 (0.280)	−0.114 (0.283)
Percent Hispanic	−3.902*** (0.993)	−2.697*** (0.848)	−2.820*** (0.885)
Median household income	−0.658*** (0.166)	−0.626*** (0.173)	−0.593*** (0.180)
Percent with limited English proficiency	−0.089 (0.078)	−0.034 (0.132)	−0.052 (0.129)
Percent unemployed	0.011 (0.164)	−0.117 (0.123)	−0.101 (0.125)
Percent with less than high school education	0.003 (0.145)	−0.046 (0.123)	−0.055 (0.122)
Observations	40,794	40,794	40,794
R2	0.946	0.947	0.947
Adjusted R2	0.941	0.941	0.941

Robust standard errors in parentheses.

^*p<0.1;

^**p<0.05;

^***p<0.01

After demonstrating that the ERPO has a statistically significant negative association with suicide mortality in implementing states’ counties, we aim to investigate whether this effect indirectly impacts counties that have not implemented the ERPO through friendship networks. To measure ERPO exposure through social networks, we defined the ERPO Social Exposure_it metric. The metric quantifies the share of the focal county’s friendship ties directed toward counties in ERPO-adopting states in year t. A formal definition of this metric is provided by equation 5 in the Methods section. We hypothesize that ERPO Social Exposure_it will be negatively associated with focal counties’ suicide mortality. This may occur because suicide-related behaviors and outcomes are interlinked with social ties. To test this hypothesis, we implement a two-way fixed-effects regression as shown in equation 6. Figure 2 provides the empirical context for the regression analysis in equation 6. The map illustrates spatial heterogeneity in the change (Δ) in ERPO Social Exposure from 2010 to 2022. Counties within states that enacted ERPO statutes during this interval (e.g., CA, WA, OR, CT, NY, NJ, MA, RI, IL, FL) exhibit elevated exposure (highlighted in yellow) due to a substantial increase in the proportion of their cross-state SCI ties with counties in other ERPO-adopting states relative to 2010. Additionally, several counties within states that did not adopt ERPO laws but share borders or strong social connections with adopting states (e.g., selected counties in NV, PA, VT) similarly demonstrate increased exposure. In these instances, we observe a rise in the exposure metric, as socially proximal states implement ERPO statutes, despite the absence of local legislative change. Consequently, the yellow shading identifies areas of indirect ERPO exposure propagated through interstate social ties, providing the necessary variation to evaluate the association between increased ERPO exposure and subsequent reductions in suicide mortality.

Figure 2: County–level change in ERPO social exposure, 2010–2022.

Colours show the change (Δ) in standardised ERPO Social Exposure between 2010 and 2022, measured in within–sample standard–deviation units. Positive (yellow) values indicate that a county’s social ties have become more concentrated in states that adopted ERPOs, whereas positive (purple) values indicate declining exposure. Although the underlying analysis covers all U.S. counties, the map shows only the 48 contiguous states and the District of Columbia; Alaska, Hawaii, and U.S. territories are not shown.

Figure 3 presents the 95% CI of the estimated coefficient ${\widehat{\delta }}_1$ associated with ERPO social exposure. Table 2, Column 2 presents estimates from the indirect-exposure specification. The ERPO Social Exposure_it coefficient is ${\widehat{\delta }}_1$ = −0.21 (cluster-robust 95% CI: [−0.34, −0.09], p < 0.01), indicating a statistically significant negative association between suicide mortality and social connectedness to ERPO-adopting states after accounting for state-by-year fixed effects (γ_st). Because ERPO Social Exposure_it is z-standardised (mean = 0, SD = 1), ${\widehat{\delta }}_1$ measures the change in suicide deaths per 100,000 population produced by a one-standard-deviation increase in out-of-state exposure. Raising a county’s exposure by one SD is associated with a reduction of 0.214 suicide deaths per 100,000. The inclusion of county-fixed effects (ϕ_i) and state-by-year fixed effects (γ_st) ensures that the estimate is identified from within-state, cross-county variation in ties to ERPO-adopting states, net of time-invariant local characteristics and contemporaneous state-level shocks. Regressions are population-weighted, and standard errors are clustered at the state level.

Figure 3:

Estimated coefficients (${\widehat{\delta }}_1$, ${\widehat{\theta }}_1$) for ERPO social exposure in two specifications. Red point indicates estimate from the baseline model without spatial exposure (${\widehat{\delta }}_1$ = −0.214, cluster-robust 95% CI: [−0.343, −0.0865]); blue point indicates estimate from the specification controlling for ERPO Spatial Exposure_it (${\widehat{\theta }}_1$ = −0.3, cluster-robust 95% CI: [−0.480, −0.121]). Horizontal lines denote 95% confidence intervals; vertical dashed line denotes the null hypothesis (δ₁ = 0). Both models include county and state-year fixed effects (ϕ_i, γ_st) and sociodemographic controls (${\overline{X}}_{\mathit{\text{it}}}$). Consistent negative and statistically significant estimates indicate the association between suicide mortality and indirect social exposure to ERPO policies is robust to spatial confounding.

To assess the robustness of this association and to differentiate the role of social networks from geographic proximity, we introduce ERPO spatial exposure. Similar to ERPO Social Exposure_it the ERPO Spatial Exposure_it metric quantifies the share of focal county’s spatial ties measured by the inverse distance between focal and alter county directed toward counties in ERPO-adopting states in year t. A formal description of this metric is shown in equation 7 in the Methods Section.

Figure 3 presents the 95% CI of the estimated coefficient ${\widehat{\theta }}_1$ as shown in equation 8. Column 3 of Table 2 presents the results estimated from two-way fixed effect regression as shown in equation 8. The coefficient on ERPO Social Exposure_it remains statistically significant at ${\widehat{\theta }}_1$ = −0.301 (cluster-robust 95% CI: [−0.48, −0.12], p < 0.01), while the coefficient on ERPO spatial exposure is ${\widehat{\theta }}_2$ = 0.53 (cluster-robust 95% CI: [−0.11, 1.16], p < 0.10).

To evaluate the robustness of our findings, we re-estimate the indirect social network exposure model using age-adjusted suicide mortality as the dependent variable. As shown in supplementary Figure S2 and Table S2, the coefficient on ERPO Social Exposure_it remains negative and statistically significant in both the indirect social network exposure model and the robust indirect social network exposure model that additionally adjusts for ERPO Spatial Exposure_it. Through these analyses, we show the existence of a negative association association between social ties and suicide-related events. This validates our hypothesis that ERPO social exposure is negatively associated with suicide mortality. We posit that the effect stems from the spread of information diffusion regarding suicidal norms, precaution, and events within social networks.

DISCUSSION

We model suicide mortality as a function of influence on socially and spatially proximate death rates, embedding counties within a socio-spatial system of interdependent risk. This framework enables the quantification of influence-induced increases in suicide rates through structured network-weighted exposure, one based on socio-spatial ties (s_−it) and the other on geographic distance (d_−it). The two-way fixed effects estimation model indicates that a one-standard-deviation increase in s_−it is associated with a statistically significant increase of 3.34 suicide deaths per 100,000 population, without controlling for spatial proximity. The directionality and magnitude of this association are consistent with research on suicide contagion in adolescent peer networks. For example, Mueller and Abrutyn [17] analyzed longitudinal friendship data from a national adolescent study and found an apparent contagion effect at the individual level. If a teenager knew that a close friend had attempted suicide, their odds of suicidal ideation were approximately 72% higher than if they did not have such exposure. Further, their odds of attempting suicide within one year were approximately 65% higher. While this study examined micro-level contagion in a youth network, our findings extend this pattern to the population level: we show the existence of associations in the suicide mortality outcomes of socially connected communities, consistent with the hypothesis of a similar contagion mechanism operating at a larger scale.

Further, when controlling for “deaths in spatial proximity” (d_−it), the estimated effect of “deaths in social proximity” (s_−it) remains statistically significant (${\widehat{\zeta }}_1$ = 2.78, cluster-robust 95% CI: [1.06, 4.50], p < 0.01). The robustness of the s_−it coefficient after controlling for deaths in spatial proximity suggests that social influence is positively associated with elevated suicide risk independent of spatial proximity. However, the reduction in the estimate of s_−it when controlling for “deaths in spatial proximity”reflects the underlying correlation between social and spatial structures, but does not eliminate the observed influence of “deaths in social proximity”. The results indicate that information about suicide related events, in this case, suicide mortality, diffuses not only along geographical proximity but also through nonlocal social ties. We speculate that this association is suggestive of a network-mediated contagion process. Our study is the first to provide evidence that county-level suicide mortality in the US exhibits socio-spatial structuring: namely, the risk of suicide death is influenced not only by physical location but also by the strength of social ties (H1).

We further extend our approach to estimate the indirect effect of a policy intervention—Extreme Risk Protection Orders (ERPOs)—via social network exposure. In a two-way fixed specification with county and state-by-year fixed effects, we find that a one-standard-deviation increase in social exposure to ERPO-adopting states is associated with a reduction of 0.21 suicides per 100,000 people (${\widehat{\delta }}_1$ = −0.21 cluster-robust 95% CI: [−0.34, −0.09], p < 0.01). This effect remains statistically significant after accounting for inverse-distance spatial exposure (${\widehat{\theta }}_1$ = −0.301, cluster-robust 95% CI: [−0.48, −0.12], p < 0.01). This,further supports the role of socio-spatial exposures. By contrast, the coefficient on spatial ERPO exposure was positive but not statistically distinguishable from zero (${\widehat{\theta }}_2$ = 0.53 cluster-robust 95% CI: [−0.11, 1.16], p < 0.10). Conditional on social exposure and state–year fixed effects, the spatial ERPO exposure term carries little independent signal. We therefore place emphasis on the robust negative association for ERPO social exposure.

These associations indicate that suicide mortality and exposure to preventive policies are patterned along measured social connectedness, beyond geographic proximity. This pattern is consistent with diffusion operating through social ties, but we do not claim that networks actively transform the content or the mechanisms of transmission. The magnitude and directionality of effects depend jointly on the structure of connectivity, the nature of the exposure, and county-level sociodemographic and economic conditions (see Table 1), which we adjust for in all models. The observed associations in our study suggest that suicide risk and its mitigation are governed by system-level processes shaped by the topology of inter-county social ties. These results indicate that suicide prevention strategies should complement geography-based targeting with decentralized, locally adaptive coordination across socially connected counties, a direction already reflected in several state policy frameworks.

Existing frameworks such as the Suicide Prevention Resource Center (SPRC) and the CDC’s Community-Based Suicide Prevention (CSP) programs already recognize the importance of cross-sectoral coordination, localized coalitions, and mental health infrastructure at the county level [38, 39]. For example, the San Diego County Suicide Prevention Council exemplifies a multi-stakeholder, network-informed approach by convening municipalities, behavioral health providers, and faith-based institutions to form an integrated response system [40].

Current prioritization schemes typically target counties based on demographic or deprivation indicators. Our results suggest complementing these approaches with SCI-derived network reach measures to identify counties with high potential to receive or transmit risk and protective influences. For instance, this can be achieved using the SCI-weighted strength of a county’s ties (the sum of its SCI weights to all other counties) or eigenvector centrality computed on the SCI matrix. Strategically positioning interventions in these counties has the potential to amplify indirect effects via social spillovers.

Network-informed infrastructures are critical for three operational reasons: (1) targeted resource deployment (e.g., gatekeeper training, telepsychiatry expansion, school-based referrals) in structurally central counties can maximize indirect impact; (2) socially connected populations enable faster diffusion of behavioral norms, including help-seeking behaviors and firearm safety practices; and (3) early warning systems built on socially connected surveillance networks can detect emerging suicide clusters before geographic clustering becomes apparent [41, 42, 43].

Limitations and Avenues for Future Research

A central component of this study is the use of the SCI to quantify inter-county social exposure to both suicide mortality and firearm-restrictive policy interventions, specifically ERPOs. SCI captures the relative probability of Facebook friendship ties between U.S. counties, offering a scalable and validated measure of population-level social connectivity [27, 44]. Prior research has shown that SCI predicts a range of real-world outcomes including migration flows, economic mobility, and infectious disease spread, underscoring its utility as a proxy for structural social networks [27, 44]. Moreover, the SCI matrix is nearly complete: all U.S. counties have non-zero SCI values, enabling comprehensive analysis of spatially distributed social influence. In our setting, this allowed estimation of suicide mortality exposure and ERPO exposure effects through non-geographic social ties, leveraging the broad geographic coverage and empirical granularity of SCI.

However, the SCI’s reliance on Facebook user data introduces several limitations. Because SCI reflects only the activity of Facebook users, it excludes individuals who are not on the platform. This results in an under-representation of older adults (65+), rural residents, and those with lower socioeconomic status or limited internet access [45]. For instance, only 50% of U.S. adults aged 65 and older report using Facebook, compared to 73% among those aged 18–29. Similarly, platform participation is lower among individuals with less than a high school education and those residing in rural areas. Consequently, SCI-derived exposure metrics may disproportionately reflect the social interactions of younger, more urban, and digitally connected populations, potentially underestimating the strength or pattern of influence in underserved subgroups. While this does not compromise the structural completeness of the SCI network, it limits the generalizability of estimated social contagion effects. Future works to address bias should consider integrating auxiliary data sources—such as the National Longitudinal Study of Adolescent to Adult Health (Add Health)—to validate findings in age- or cohort-specific networks. Furthermore, methodological innovations, including dynamic or data-driven influence weighting (e.g., via latent space models or dynamic Bayesian kernels), may allow more flexible modeling of inter-county exposure beyond the fixed SCI structure.

While our current specification defines the social proximity weights w_ij through a parametric kernel based on SCI and population scaling, this structure imposes a fixed functional form on how social influence decays across counties. Although the current model incorporates a dynamic specification for social exposure by adjusting the strength of social influence based on suicide rates, the overall structure of the social proximity weights remains pre-specified and fixed. This parametric formulation is transparent and computationally efficient but may restrict the model’s ability to reflect the true heterogeneity of influence processes, particularly in settings where socially central counties exert disproportionate or nonlinear effects on others. In future work, these limitations could be addressed by adopting a fully Bayesian hierarchical framework in which the social influence weights are not assumed a priori, but instead are learned from the data. For example, one could introduce latent interaction structures or flexible prior distributions that allow the strength and topology of social exposure to vary across space and time. This would permit more granular and context-sensitive modeling of how influence propagates through the social network. Such an approach would enable the model to capture emergent patterns of contagion or resistance that may be overlooked under fixed-weight assumptions. Advancing toward data-driven, nonparametric, or semi-parametric Bayesian inference represents a rigorous path for improving the fidelity and generalizability of network-based models of suicide mortality in population health research.

MATERIALS AND METHODS

This study integrates multiple county level datasets to construct time-varying measures of suicide mortality, socio-spatial influence, spatial influence, social exposure, spatial exposures and county-level covariates.

Suicide mortality

Data on suicide deaths were obtained from the National Vital Statistics System (NVSS), managed by the National Center for Health Statistics (NCHS). We extracted yearly counts of county-level mortality from 2010 to 2022 using the International Classification of Diseases, Tenth Revision (ICD-10) codes: X60–X84, Y87.0, which are conventionally used to identify deaths due to intentional self-harm. For confidentiality and consistency, all death counts were aggregated to the annual level and standardized per 100,000 population using county-level denominators from the U.S. Census Bureau.

Social connectedness index

The Social Connectedness Index (SCI) was obtained from the Meta Data for Good program. SCI quantifies the relative strength of Facebook friendship ties between counties, normalized by population size. Specifically, SCI_ij reflects the likelihood of a Facebook user in county i being friends with a user in county j. This measure provides a high-resolution, empirical proxy for social ties throughout US counties. Formally, the SCI between two locations i and j is defined as follows [27]:

$${\text{SCI}}_{\mathit{\text{ij}}}=\frac{{\text{Facebook}\text{Connections}}_{\mathit{\text{ij}}}}{{\text{Facebook}\text{Users}}_i\times {\text{Facebook}\text{Users}}_j}.$$

Here, Facebook Users_i represents the number of Facebook users in the county i. Facebook Connections_ij is the total number of Facebook friendship connections between individuals in counties i and j. Data were accessed for the year 2020 and assumed temporally stable over the study period.

Socioeconomic covariates

Time-varying county-level sociodemographic characteristics were drawn from two sources. First, the Agency for Healthcare Research and Quality (AHRQ) provides cross-sectional community-level indicators, including education, unemployment, income, and racial/ethnic composition, for 2010–2020. Second, for years not covered by AHRQ, we accessed the American Community Survey (ACS) via the tidycensus package in R, using 5-year estimates for all U.S. counties. The harmonized covariates include population density, age structure, median income, educational attainment, unemployment rate, and proportion of residents with limited English proficiency.

Socio-spatial influence metrics

To quantify the influence of suicide mortality in socio-spatially connected counties, we construct two metrics: social proximity to suicide deaths (s_−it) and spatial proximity to suicide deaths (d_−it). These are formally defined as:

$$s_{-\mathit{\text{it}}}=\sum _{j\ne i}{w}_{\mathit{\text{ij}}}{y}_{\mathit{\text{jt}}},d_{-\mathit{\text{it}}}=\sum _{j\ne i}{a}_{\mathit{\text{ij}}}{y}_{\mathit{\text{jt}}}$$ (1)

where y_jt denotes the suicide death rate in county j at time t. The weights w_ij and a_ij represent social and spatial proximity, respectively, defined as:

$$w_{ij}=\frac{n_j\cdot {\text{SCI}}_{\mathit{\text{ij}}}}{{\sum }_{k\ne i}{n}_k\cdot {\text{SCI}}_{\mathit{\text{ik}}}},a_{\mathit{\text{ij}}}=\frac{1/d_{\mathit{\text{ij}}}}{{\sum }_{k\ne i}\left(1/d_{\mathit{\text{ik}}}\right)}$$ (2)

where n_j is the population of county j, SCI_ij is the Social Connectedness Index between counties i and j, and d_ij is the great-circle distance between them. Consistent with prior literature [44], we account for correlation between these metrics due to the spatial clustering of social networks by including both terms jointly in the model specification.

To examine the association between socio-spatial influence and suicide mortality, we estimate the following two-way fixed effects regression model:

$$y_{\mathit{\text{it}}}={\zeta }_1{s}_{-\mathit{\text{it}}}+{\zeta }_2{d}_{-\mathit{\text{it}}}+{\overline{\zeta }}_3^T{\overline{X}}_{\mathit{\text{it}}}+{\mu }_i+{\varphi }_t+{\epsilon }_{\mathit{\text{it}}},$$ (3)

where y_it is the suicide death rate in county i and year t; ${\overline{X}}_{\mathit{\text{it}}}$ is a vector of time-varying sociodemographic covariates (e.g., age distribution, racial/ethnic composition, income, education, unemployment, limited English proficiency); μ_i and ϕ_t are county and year fixed effects, respectively. Estimation is conducted using population-weighted least squares, and standard errors are clustered at the state level to account for within-state correlation in the error structure.

ERPO policy social and spatial exposure metric

We estimate a two-way fixed effects model to assess the direct influence of Extreme Risk Protection Order (ERPO) policies on suicide mortality at the U.S. county level. The baseline specification is:

$$y_{\mathit{\text{it}}}={\psi }_1\mathit{\text{ERP}}{O}_{\mathit{\text{it}}}+{\overline{\psi }}_{\text{2}}^T{\overline{X}}_{\mathit{\text{it}}}+{\varphi }_i+{\gamma }_t+{\epsilon }_{\mathit{\text{it}}},$$ (4)

where y_it denotes the suicide death rate per 100,000 population in county i and year t, and ERPO_it is a binary indicator equal to 1 if the ERPO policy was enacted in state s(i) at time t. The vector ${\overline{X}}_{\mathit{\text{it}}}$ includes socioeconomic and demographic covariates (e.g., age distribution, racial/ethnic composition, income, education, unemployment, limited English proficiency). County (ϕ_i) and year (γ_t) fixed effects control for unobserved, time-invariant heterogeneity and national temporal shocks, respectively. We exclude state-by-year fixed effects from this model, given that ERPO_it varies at the state-year level. Estimation is performed using population-weighted ordinary least squares, and standard errors are clustered at the state level.

To quantify indirect influence transmitted through social networks, we define the following metric:

$$\mathit{\text{ERPO}}\mathit{\text{Social}}\mathit{\text{Exposur}}{e}_{\mathit{\text{it}}}=\sum _{s(i)\ne s(j)}^S\mathbb{1}{\left(\text{ERPO}\text{in}\text{state}s\left(j\right)\right)}_t\times \frac{\mathit{\text{SC}}{I}_{\mathit{\text{ij}}}}{{\sum }_h\mathit{\text{SC}}{I}_{\mathit{\text{ih}}}}$$ (5)

where SCI_ij denotes the Social Connectedness Index between counties i and j, and the indicator function reflects ERPO policy adoption in other states s(j) ≠ s(i). This standardized metric captures the share of social ties that county i maintains with counties located in ERPO-adopting states.

To estimate the association between social exposure and suicide mortality, we specify:

$$y_{\mathit{\text{it}}}={\delta }_1\mathit{\text{ERPO}}\mathit{\text{Social}}\mathit{\text{Exposur}}{e}_{\mathit{\text{it}}}+{\overline{\delta }}_3^T{\overline{X}}_{\mathit{\text{it}}}+{\varphi }_i+{\gamma }_{\mathit{\text{st}}}+{\epsilon }_{\mathit{\text{it}}},$$ (6)

where ϕ_i and γ_st represent county and state-by-year fixed effects, respectively. This specification accounts for unobserved heterogeneity across counties and time-varying factors at the state level, including a county’s own state-level ERPO implementation. The model is estimated using population-weighted OLS, with standard errors clustered at the state level to allow for arbitrary correlation within states. The inclusion of state-by-year fixed effects ensures that estimated effects are identified from variation in out-of-state ERPO policies among socially connected counties, while controlling for time-varying enforcement capacity, socioeconomic context, and other latent shocks at the state level.

As a robustness check, we define a spatial exposure metric that captures geographical proximity to ERPO implementation in neighboring states:

$$\mathit{\text{ERPO}}\mathit{\text{Spatial}}\mathit{\text{Exposur}}{e}_{\mathit{\text{it}}}=\sum _{s(i)\ne s(j)}^S\mathbb{1}{(\text{ERPO}\text{in}\text{state}s(j))}_t\times \left(\frac{1/d_{\mathit{\text{ij}}}}{{\sum }_{k\ne i}\left(1/d_{\mathit{\text{ik}}}\right)}\right)$$ (7)

where d_ij denotes the great-circle distance between counties i and j, excluding counties in the same state s(i). This formulation captures whether ERPO implementation in geographically proximate counties affects suicide mortality via spatial diffusion.

To evaluate the joint influence of both social and spatial spillovers, we estimate the following model:

$$y_{\mathit{\text{it}}}={\theta }_1\mathit{\text{ERPO}}\mathit{\text{Social}}\mathit{\text{Exposur}}{e}_{\mathit{\text{it}}}+{\theta }_2\mathit{\text{ERPO}}\mathit{\text{Spatial}}\mathit{\text{Exposur}}{e}_{\mathit{\text{it}}}+{\overline{\theta }}_3^T{\overline{X}}_{\mathit{\text{it}}}+{\varphi }_i+{\gamma }_{\mathit{\text{st}}}+{\epsilon }_{\mathit{\text{it}}},$$ (8)

where ϕ_i and γ_st denote county and state-by-year fixed effects, and ${\overline{X}}_{\mathit{\text{it}}}$ represents time-varying covariates. This model allows for simultaneous estimation of the independent associations between suicide mortality and social vs. spatial channels of ERPO influence. Population-weighted estimation and state-level clustered standard errors ensure inference is robust to differential population sizes and intra-state dependence in error terms.

Code Availability

Analysis code to reproduce figures and tables in the paper is available at https://github.com/kut97/suicide-sci

Data Availability

Mortality data was obtained from NVSS, managed by the National Center for Health Statistics (NCHS) [46]. Due to confidentiality concerns, this data set is not publicly accessible, but can be requested from NCHS at https://www.cdc.gov/nchs/nvss/nvss-restricted-data.htm. Social determinants of health (SDOH) covariates for 2010–2020 were obtained from the Agency for Healthcare Research and Quality (AHRQ) SDOH Database (https://www.ahrq.gov/sdoh/data-analytics/sdoh-data.html); for 2020–2022, SDOH covariates were constructed from U.S. Census Bureau products (American Community Survey [ACS] 5-year estimates; see https://www.census.gov/programs-surveys/acs/data.html). The Social Connectedness Index (SCI) is available through Facebook (Meta) Data for Good at https://dataforgood.facebook.com/dfg/tools/social-connectedness-index. For age-adjusted suicide mortality, annual population denominators stratified by 18 five-year age groups were drawn from CDC WONDER Bridged-Race Postcensal Population Estimates (https://wonder.cdc.gov/single-race-population.html).