Suicide is a complex issue with risks occurring at the individual, relationship, community, and societal levels (U.S. Department of Health and Human Services, 2012; Stone et al., 2018; Virginia Department of Health, 2016). Suicide risk varies according to age, sex, race, and other demographic factors. Some common circumstances preceding a suicide death include mental health problems, relationship problems, a recent crisis, alcohol or substance misuse, physical health problems, and financial problems. While understanding individual-level risk factors is essential for suicide prevention efforts, exploring spatial patterns and identifying high-risk areas of suicide can inform more targeted and comprehensive prevention efforts and improve resource allocation. Culturally appropriate suicide prevention interventions that address specific risk factors in different populations and places are most effective (Barnhorst et al., 2021).

There are some important limitations to methods that have been previously used to identify geographic clusters of suicide. Several studies have used spatial scan statistics to identify geographic clusters of suicide and characteristics associated with the clusters (Fontanella et al., 2018; Kulldorff & Nagarwalla, 1995; Saman et al., 2012); however, this method will identify the most likely clusters, even if they are not significantly different from the rest of the study area. Bayesian spatial regression is another method for identifying geographic clusters of suicide. One such study identified 52 counties in Virginia with greater than expected suicide risk and found that suicide risk was positively associated with the percentage of the White population and higher median age (Orndahl & Wheeler, 2018). However, using county boundaries severely limits the ability to detect geographic clusters since the risk of suicide can be highly concentrated in only one part of a county and can cross county boundaries. Therefore, new methods for identifying small-area geographic suicide clusters and community-level risk factors should be explored.

An important aspect of our method for detecting geographic clustering of suicide risk is our use of the National Violent Death Reporting System (NVDRS). The NVDRS is a state-based surveillance system containing individual-level data about each suicide, including demographic characteristics and the residential tract of each decedent. Importantly, this surveillance system contains unique information about individual circumstances that precipitated a suicide. Our method for detecting geographic clusters of high suicide risk differs from other studies because all geographic units in the state have a uniform and minimum level of statistical reliability instead of a minimum level of geographic precision. We accomplished this using a series of overlapping moving windows called spatially adaptive filters (Cai et al., 2011; Talbot et al., 2000; Tiwari & Rushton, 2005). Spatially adaptive filters are aggregations of smaller neighboring geographic units (in this case, Census tracts) that, by themselves, do not have sufficiently large populations to calculate statistically reliable disease rates (Matthews, 2018). The size of the spatial filters varies according to population density; filters are smaller in urban areas and larger in rural areas. Others have used spatially adaptive filters to create an interpolated map of disease rates with a uniform statistical reliability for other diseases. However, identifying geographic clusters, areas where disease rates are statistically significantly elevated compared to the state overall, is a novel application of spatially adaptive filters. Using Virginia as an example, we identified clusters with elevated suicide rates and then compared the suicide circumstances of the decedents residing within the clusters to all other parts of the state outside the clusters.

We analyzed data for 6,218 suicide deaths reported to Virginia VDRS from 2010 through 2015. We identified 13 high-risk suicide clusters, which captured 1,005 (16.1%) suicides in the state over the six years. These high-risk clusters accounted for 8.7 % (n = 166) of the Census tracts in Virginia and represented 9.0% (n = 626,864) of the population at-risk. The clusters were dispersed throughout the state and the geographic variation in the IAR is high (Figure 1). The clusters had a population ranging between 22,915 and 124,232 and an SMR for suicide ranging between 1.7 and 2.0 (Table 1). Six of the clusters contained rural Census tracts as defined by RUCC.

The clusters differed from the rest of the state for certain suicide decedent characteristics (Table 2). Suicide decedents in the clusters were more likely to be older (40+ years), White, and widowed/divorced/separated than decedents in the rest of the state (i.e., decedents outside the clusters). Firearm was the most common suicide method in clusters and the rest of the state, accounting for 62% of suicides in clusters compared to 55% of suicides in the rest of the state. The proportion of suicide by firearm was significantly higher in cluster 9 (78%, P < .01), cluster 11 (79%, P < .01), cluster 12 (72%, P < .01), and cluster 13 (75%, P < .01) compared to the rest of the state. The proportion of suicide by poisoning was significantly greater in clusters 1 (25%, P < .01) and 2 (29%, P < .01) compared to the rest of the state; however, the proportion of suicide by poisoning was not significantly different from the rest of the state for all clusters combined.

We found differences in decedents’ suicide circumstances between clusters (Table 2); however, no precipitating suicide circumstance was more prevalent for all clusters combined compared to the rest of the state. “Current mental health problem” was the most common circumstance in all clusters (58%), followed by “history of mental illness treatment” (49%) and “current mental illness treatment” (39%). Some individual clusters differed from the rest of the state for specific suicide circumstances. Clusters 4, 7, and 11 had significantly lower proportions of suicides with reported mental health problems than the rest of the state (Cluster 4: 38%, P = .01; cluster 7: 39%, P < .01; cluster 11: 40%, P = .02). Similarly, the proportion of suicides with current mental illness treatment was lower in cluster 6 (28%, P = .04) and the proportion of suicides with a history of mental illness was lower in cluster 7 (37%, P = .03) compared to the rest of the state. The proportion of suicides with “job problems” as a precipitating circumstance was significantly lower in clusters 9 through 13 (range: 2–8%; P ≤ .02) compared to the rest of the state. The proportion of suicides with “financial problems” was significantly lower in cluster 12 (5%, P < .01) and cluster 13 (4%, P < .01) compared to the rest of the state; similarly, the proportion of suicides with “eviction or loss of home” was significantly lower in cluster 12 (1%, P < .01) and cluster 13 (0%, P = .03). Among all clusters combined, the proportion of suicides with job problems (11%, P < .01), financial problems (12%, P = .02), and eviction or loss of home (3%, P = .04) was significantly lower than the rest of the state.

This analysis demonstrates the potential utility of enhancing surveillance systems such as NVDRS with small-area level geographic data. The pairing of geographic information with surveillance data can assist in the identification of both areas with higher than state average suicide rates and place-specific suicide risk factors. In this descriptive analysis, we described the location of high-risk areas in the state to encourage future investigations into causes and protective factors of suicide in Virginia and to develop data-driven, targeted suicide prevention activities.

We used a novel approach to identify clusters with spatially adaptive filters, which diverges from the contemporary literature on suicide cluster identification. The most commonly used method for detecting suicide clusters, the spatial scan statistic method, identifies a most likely cluster even when the statistical significance of the test statistic is low. However, our analysis used spatially adaptive filters as the basis for our clustering method to address the impact that different population sizes have on the statistical reliability of the disease rates (Choynowski, 1959; Waller et al., 2006). While other studies have used spatially adaptive filters to represent geographic patterns of disease rates as a continuous surface (Figure 1B), we extended the use of spatial filters as a new way to identify geographic clusters. In doing so, we detected several highly geographically detailed clusters where suicide rates were significantly higher than in Virginia (Figure 1A). Moreover, the identified clusters in this analysis tended not to follow county administrative boundaries; they either occurred within counties or contained regions from neighboring counties. These results could inform future work examining sub-county clustering of suicide and changes in suicide clusters over time.

This analysis revealed important differences in suicide methods between high suicide risk clusters. Compared to the rest of the state, four clusters in western Virginia had a significantly higher proportion of suicides from firearm-related injuries, and two clusters in northern Virginia had a significantly higher proportion of suicides from poisoning. These clusters contain a higher proportion of rural Census tracts than any other cluster in the study. An important driver of urban-rural differences in suicide is the increased rate of suicide by firearm in rural areas (Ivey-Stephenson et al., 2017; Nestadt et al., 2017). Two firearm suicide clusters were previously identified in Ohio, in the Appalachian region of the state (Fontanella et al., 2018); the clusters with a higher proportion of firearm suicides in this paper, which also occurred in or near the Appalachian region of Virginia, may indicate larger regional trend. Poisoning has a relatively low case fatality rate, which may suggest high levels of non-fatal substance misuse in the clusters with a higher proportion of poisoning suicides (Miller et al., 2004).

The pattern of mental health circumstances in all suicide clusters combined was not different from that of the rest of the state, although the proportion of decedents reporting mental health circumstances did differ for some individual clusters. Overall, mental health circumstances were common among decedents inside and outside clusters, which underscored the importance of preventing and treating mental health conditions for suicide prevention. However, some individual clusters reported a significantly lower proportion of mental health conditions (clusters 4, 7, & 11) and mental health treatment (cluster 6) compared to the rest of the state. Treatment for mental health conditions could be affected by various individual (e.g., health insurance status, mental health condition) and environmental (e.g., health and mental health provider density) factors. Furthermore, these results do not account for regional variations in mental health care, such as differences in quality of care between urban and rural areas (Gamm et al., 2010; Ziller et al., 2010).

Job problems, financial problems, and eviction or loss of home were less likely to be reported as precipitating suicide circumstances in all clusters combined compared to the rest of the state, although these circumstances varied regionally. Decedents in three high-risk clusters (9–11) in western Virginia were less likely to have known job problems, and decedents in two high-risk suicide clusters (12 & 13) in rural Appalachian Virginia were less likely to have known job problems, financial problems, and eviction or loss of home compared to the rest of the state. Some research has found an association between individual socioeconomic disadvantage and suicide, but the association is inconsistent (Burrows et al., 2011). The results from this analysis may indicate the relative importance of precipitating factors other than job problems, financial problems, and eviction for suicide in clusters 9–13.

Information about spatial variation in suicide rates could help direct suicide prevention resources to areas with the greatest need in Virginia and elsewhere. These data could encourage the development of more targeted, effective prevention programs, such as strategies described in the CDC’s suicide prevention technical package (Stone et al., 2017). The integration of small-area geographic data to NVDRS provides valuable information about spatial variation in suicide risk factors that can facilitate place-based suicide prevention strategies and be used in small-area geographic analyses with other topics (e.g., homicide). This analytic strategy is useful for guiding targeted suicide prevention efforts and informing additional research to understand the increasing rates of suicide.