An exploratory study on spatiotemporal clustering of suicide in Korean adolescents

This study used data from student suicide reports collected by the Korean Ministry of Education from January 1, 2016, to December 31, 2020. In Korea, when a student dies by suicide, the school is required to report the relevant information to the Ministry of Education in the student suicide report, which includes teachers’ observations, parental reports regarding the circumstances of death, and official education records collected by the school. Furthermore, these reports were collated as part of the national student suicide prevention policy during the abovementioned period. The evaluation items and answer format were determined through intensive discussion within the research team and feedback from teachers during the report’s development process. Additionally, specific examples of items and answers were provided in the form to simplify it and enable the teachers to understand and respond better. During the coding process, unclear answers were deciphered through discussion within the research team and confirmed by contacting the teacher directly [37, 38]. These data represent the total number of students who died by suicide in Korea during the study period. Details of the student suicide reports have been described previously [38]. The number of students who died by suicide during the study period was 654, and all cases were included in the analyses except for two students whose death dates could not be determined. Considering that Korea has compulsory education up to middle school and the dropout rate of high school in 2021 is 1.5% [39], these cases may closely represent the general characteristics of suicides among children and adolescents in Korea.

The variables used in this study were the address of the school, sex, date of death, school type, family structure, economic status, suicide method, usual concerns revealed at school, presence of a psychiatric disorder, history of suicide attempt, and history of self-injury. The teacher-rated Strengths and Difficulties Questionnaire (SDQ) [40] was used to evaluate students' emotional and behavioral status. The teacher-rated SDQ consists of Prosocial Behavior (Cronbach’s α = 0.873), Hyperactivity/Inattention (Cronbach’s α = 0.793), Peer Relationship Problems (Cronbach’s α = 0.770), Emotional Symptoms (Cronbach’s α = 0.681), Conduct Problems (Cronbach’s α = 0.638) subscales and a Total Difficulties score (Cronbach’s α = 0.837). The SDQ has been included in the database since 2018. This study was approved by the Institutional Review Board of Hallym University Sacred Heart Hospital (2021-05-015).

The school addresses of students who died by suicide were converted to latitude and longitude coordinates to examine the proximity of both the space and time of suicidal events, with the time of occurrence set on the day of the event. For cases with incomplete information regarding the date of death, information on the time of discovery was used. As approximately 70% of cases of adolescent suicide in Korea die by jumping from a height, the interval between the time of a suicide attempt and the time of death was expected to be short.

Clustering analysis using density-based spatial clustering of applications with noise (DBSCAN) [33, 41] was used to examine the spatiotemporal patterns of suicidal events and define the space–time clusters of suicides. Density-based clustering refers to unsupervised learning methods that identify distinctive groups or clusters in the data based on the idea that a cluster in a data space is a contiguous region of high point density, separated from other clusters by contiguous regions of low point density. The data points in the separating regions of low point density are typically considered noise/outliers [33, 41]. In particular, this method is useful when there is an outlier in the spatial information that is included in a cluster and distorted [42].

The two main conditions to be considered in DBSCAN for the derivation of clusters are the minimum number of cases to be included in the cluster and the cluster radius. In this study, the minimum number of suicide clusters was set at three. The radius of the cluster was selected by examining the change in the distance of the k-nearest neighborhood (kNN). The k value was set to three to simulate the kNN point change, which was equal to the minimum number of clustering cases. R version 4.2.2 was used for analysis and the cluster analysis was performed using the R language DBSCAN package (Hahsler et al.). The proximity among cases within the cluster is represented by the mean distance (mdis), where a lower numerical value indicates closer clustering of cases.

Figure 1 presents the spatiotemporal distribution of the suicide cases. Figure 1a shows the distribution of suicide case events on the map of South Korea, and the year of the event is also marked in a different color. Many cases were distributed around large cities with dense populations. However, even in areas with relatively sparse populations, suicide cases occur at a certain level. Figure 1b shows the results of standardizing the latitude, longitude, and time to place the case in 3-dimensional space and demonstrates that the distribution of suicide events does not occur randomly but rather clusters in a specific space–time area.

Figure 2a presents the results of the analyses that examined the change in the distance of the 3-nearest neighborhood to determine the criterion of the radius of the cluster prior to DBSCAN. In the figure, the knee appears around the distance of 60. Figure 2b presents the clustering results when the radius was set to 60 and the minimum number of cases belonging to a cluster was set to three. Each cluster is presented as a polygon. Outliers that did not belong to any cluster were marked as separate dots. As shown in the figure, the size of the cluster and the number of included cases varied. The largest cluster at the top of the figure reflects spatially concentrated suicide cases in densely populated areas in the Seoul metropolitan area. However, these suicides demonstrated a wide temporal distribution spanning approximately 4 years.

Table 1 presents the characteristics of the clusters derived using DBSCAN. Along with the closeness of the cluster (mdis) and number of cases in each cluster’s data distribution, the table also shows the proportion of males, high school students, middle school students, and mean age. Next, the first occurrence date, last occurrence date, longitude, and latitude of the schools attended by the students who died by suicide are presented. The latitude and longitude of the clusters were determined using the average latitude and longitude of the schools within the clusters. Significant clusters are listed in order of the smallest mdis size. Finally, they are listed based on the size of the clusters. For example, in the case of Cluster 1, which is the cluster with the most substantial spatiotemporal proximity, five cases of suicide centered on a specific area occurred within approximately 3 weeks. All patients in this cluster were high school students, and all but one were adolescent boys.

We identified 23 clusters through data analysis of 652 cases using DBSCAN. The largest cluster (class ID = 23) comprised of 395 patients. The period of the events covered approximately 5 years. As mentioned above, this cluster could result from demographic concentration, especially in the context of urban South Korea with high population density, rather than from space–time suicide clusters. Therefore, defining a significant suicide cluster that shows a remarkably high spatiotemporal adjacency.

Figure 3 presents the results of comparing the cluster closeness (mdis) and the number of cases in the cluster data. The ranking on the horizontal axis is the result of sorting by area. The upper part of Fig. 3 presents all the clusters, and the lower part shows the figure, excluding the largest cluster. As shown in the figure, the area and number of cases rapidly increased after the 15th cluster. Based on this finding, the meaningful spatiotemporal cluster of suicide was defined as up to the 15th cluster (class ID = 12) based on the rank number. We identified 63 (9.7%) spatiotemporally clustered suicides among adolescents, with a temporal range between 7 and 59 days. In the case of spatial range, each cluster was analyzed in a polygonal form, making it difficult to precisely ascertain the average spatial area. Nonetheless, cases classified into significant clusters were predominantly within the same administrative regions. When considering the top three clusters with the highest spatiotemporal clustering (Ranks 1–3 in Table 1), the closest distance between the two suicide cases was approximately 6 km, and the greatest distance observed was approximately 32 km.

Table 2 shows the comparison of the characteristics of a group that showed high spatiotemporal clustering in suicide with those of a group that did not. Chi-square analysis revealed that the characteristic that was statistically different between the two groups was economic status (χ² = 9.79, df = 2, p < 0.05). The clustered suicide group was relatively low. Although no difference was observed at the stochastic significance level, participants showing clustered groupness were relatively more likely to experience peer problems. In the group without significant spatiotemporally clustered groupness, 15.1% (n = 89) reported peer problems, and in the group with clustering, 23.0% (n = 15) reported problems with peer relationships. The reported rate of psychiatric disorders was 29.7% (n = 19) in the clustering group and 40.8% (n = 231) in the other groups.

Table 3 presents the results of the group comparisons using the SDQ. The results of the t-tests indicated that there were no statistically significant differences between the two groups for SDQ total and subscale scores.

In Table 4, the binary logistic regression analysis results are presented, with the highly clustered group being the outcome variable and the non-clustered cases being the reference group. The demographic and clinical characteristics that were found to significantly differ based on group included economic status (e.g., poverty) and the presence of a psychiatric disorder (p < 0.05). As the economic level decreases (indicative of poverty), there is an increased tendency for spatiotemporal clustering. However, the less likely the cases included reported psychiatric disorders, the more likely they were to be in a highly clustered group. Groups that reported peer problems had a higher likelihood of being highly clustered, even though the statistical significance of this result was low (p < 0.10).

In this study, 9.7% of adolescent suicides were classified into spatiotemporal suicide clusters, which was a higher percentage than previously reported. This increase could be attributed to differences in the analytic methods. The current findings suggest that interrelated suicides may be more frequent than expected in adolescents [5]. However, this study statistically identified spatiotemporal suicide clusters but did not confirm that suicides within clusters were actually related to suicides. Suicides that occurred within similar time periods in similar locations could have been classified into this space–time suicide cluster, even if there was no real connection. Future research should include a detailed case study of the suicide cases in these clusters.

The mechanisms leading to suicide clusters include social transmission, particularly person-to-person transmission and the media [5, 6, 18]. In addition, clustered suicide occurs through perceptions that suicidal behavior is widespread and assortative, leading to susceptible young people being likely to socialize with at-risk peers, and the social cohesion of the peer group contributes to the spread of ideas and attitudes [5, 6]. The effect of suicide clusters on schools is usually profound, and the early identification of suicide clusters and initiation of appropriate interventions is critical for preventing subsequent suicides. This study suggests that once an adolescent died by suicide, close monitoring and intervention may be needed to prevent subsequent suicides for about 2 months.

Several features of the clustered adolescent suicides in this study were similar to the socioeconomic characteristics of previously identified clustered suicides. Previous studies have identified deprivation [31, 43], poverty [22, 23, 43‐45], and geological isolation [25, 30, 46] as significant risk factors for clustered suicide. In this study, economic status was lower among the clustered suicides than the non-clustered suicides in both the chi-square test and logistic regression analysis, which mirrors the results of previous studies.

In previous studies, young men were more frequently included in the clustered suicide groups than were young women [24, 47, 48]. However, this finding has not been replicated in other population-based studies targeting young adults and adolescents [5, 11, 25]. Similarly, there was no difference in the gender ratio between clustered and non-clustered suicides in our study, the first to report the gender characteristics of clustered suicides in Korean adolescents. This could be due to differences in the analytical method (DBSCAN) used to identify suicide clusters between this study and previous studies. Furthermore, 654 suicides were included in the analysis, which is fewer than in previous studies; this could have potentially influenced the results. Hence, future studies that target a larger number of suicides over an extended period are needed.

Regression analysis revealed that the clustered suicide groups had fewer psychiatric disorders than the non-clustered suicide group. This differs from previous findings and suggests that psychiatric history is a risk factor for clustered suicides [6]. However, it should be noted that the assessment of psychiatric disorders among students who died by suicide was based on parental reports after suicide rather than the direct application of standardized diagnostic tools, thus potentially failing to adequately capture the frequency of psychiatric disorders. Even if the students had clinically diagnosed psychiatric disorders, they may not have visited hospitals because of negative perceptions associated with mental health or that parents did not accurately report due to concerns about potential disadvantages the students might face at school. Additionally, no statistically significant differences were observed regarding the presence of psychiatric disorders between the two groups in the chi-square test. Given the limited number of participants, further research is necessary to address these findings.

Another distinctive characteristic of clustered group was their low economic status, which is consistent with previous studies [18, 22, 23, 31, 43‐45]. However, earlier studies have not clarified the relationship between socioeconomic status and suicide clustering. In some studies [22], low economic status has been suggested as a proxy for factors associated with the clustering of suicides, such as limited access to mental health treatment. Since limited information was collected from each participant, our study could not clearly explain the underlying mechanism. Considering the multidimensional risk factors of adolescent suicide [49], and the general social stigma against psychiatric disorders in South Korea [50], having a low economic status might also decrease help-seeking behavior for the early detection of mental health problems of clustered suicide adolescents in Korea.

In summary, by using DBSCAN to analyze clustered adolescent suicides in Korea, we found a higher rate (9.7%) than that reported in previous studies. Moreover, the temporal range for the clustered suicides identified was within 2 months. These suicides were characterized by lower economic status, which is consistent with previous studies [22, 23, 43‐45]. Our study differs from previous studies in that we used a methodology that did not use a specific window, providing a basis for identifying the critical time and regions for subsequent adolescent suicide prevention.

This study has several limitations. First, suicide cases among adolescents used in our study only included those reported by schools; thus, out-of-school adolescents were excluded. Second, our study exclusively focused on Korean students who died by suicide over 5 years, resulting in a limited sample size. This is because our study was a secondary analysis of data collected during a limited period, 2016–2020, as part of a suicide prevention policy in Korea. Correspondingly, given the exclusive focus on Korean adolescents, the distinct attributes of suicide may be influenced by national and cultural contexts, impeding the generalization of this study’s outcomes to diverse international settings. Third, we defined clustering as involving a minimum of three suicides; thus, cases in which two consecutive suicides occurred in a spatiotemporal context similar to clustered suicides were not included. Fourth, we did not account for factors that could link adolescents who died by suicide, even when not in geographically similar spaces, such as the Internet or social network services. Therefore, clustered suicides among adolescents might not have been adequately identified. Finally, the geographic data employed in this study were derived from school addresses rather than the residential addresses of adolescents who died by suicide, consequently failing to accurately reflect the specific locations of suicide incidents. However, Korean students are assigned to schools through a system known as the school district [51], wherein the proximity of a student’s residence serves as the paramount criterion for school assignment. Therefore, the addresses of the schools utilized in our research can indirectly represent the actual places of residence, and this window is much narrower than the previous studies that used county-level data. Additionally, in the Korean context, the living environment and peer groups of adolescents are often organized on a school-based scale, thereby highlighting the significance of the findings of this study.