Adolescents’ suicidal thoughts and behaviors are pressing global concerns and have been on the rise in recent years. Despite extensive research, progress in comprehensively understanding the risk and protective factors associated with suicidal thoughts and behaviors remains limited. Conceptually, the relationship between deliberate, planned self-destructive behaviors in adolescents, such as suicide attempts, and more impulsive or unintentional forms of selfdestruction, like eating disorders and life-threatening risk-taking, remains both unclear and insufficiently studied. Methodologically, uncertainties persist about how to study complex phenomena such as suicidal ideation or suicide attempts holistically, considering the interplay of a wide range of factors across various domains. Consequently, empirical research and theoretical frameworks have primarily emphasized a limited set of factors as key drivers of suicidal thoughts and behaviors, often overlooking a range of other biopsychosocial domains and cognitive processes that may contribute to lethal forms of both intentional and unintentional selfdestruction.
This thesis addresses these limitations by proposing a holistic multimodal domain analysis (HMDA) approach. This approach assesses a broad spectrum of risk and protective factors in a complex system and identifies the significant factors in the system, while taking their interactions into account. Moreover, it employs hundreds of machine learning models and considers the disagreements between them, providing a robust and comprehensive understanding of key predictors, deduced factors, and the reliability of the assessment. The thesis employs this approach to explore multiple domains and identify factors associated with suicidal behavior as well as to assess the reliability of self-reported suicide attempts. It also proposes a new theoretical framework for understanding suicidal tendencies and different forms of self-destructive behaviors in adolescents. Furthermore, the thesis validates and compares the proposed theory with three wellestablished theories of suicide—the Interpersonal Theory of Suicide (Joiner, 2005), the Three-Step Theory of Suicide (Klonsky & May, 2015), and the Strain Theory of Suicide (Zhang, 2016)—using nationally representative samples of Norwegian adolescents.
Research questions
This thesis addresses several core research questions: 1) What are the important risk and protective factors associated with adolescent suicide attempts? 2) Which theory of suicide is most compatible with the identified risk and protective factors? 3) Is it possible to identify adolescents at risk of attempting suicide without relying on any suicide-related information such as suicidal ideation, suicide plans, nonsuicidal self-harm, or fearlessness about death? 4) Is it possible to predict the future onset of suicide attempts in adolescents? 5) What factors predict future inconsistencies in self reports of lifetime suicide attempts?
Methods
Given the low prevalence of suicide attempts among adolescents, reliable empirical research requires large sample sizes. This thesis utilized two extensive datasets: one cross-sectional dataset (collected between 2014 and 2019) and one longitudinal dataset (collected between 1992 and 1994). The time gap between these datasets offered the opportunity to assess risk and protective factors across different generations of Norwegian adolescents and assess the validity of the aforementioned theories for both samples. All datasets were holistically analyzed at the item level using HMDA, although this approach gradually matured during the study period. This thesis also makes other significant methodological contributions. It introduces an innovative machine learning solution for missing data imputation of nested mixed-type survey data (e.g., psychological surveys collected in schools), which automatically handles imputing imbalanced (low-prevalence) items. Additionally, a new algorithm specifically designed for modeling low-prevalence outcomes is presented. These algorithms were developed in response to methodological obstacles encountered during the PhD period and their importance in advancing research methods for
studying suicidal behavior is emphasized.
Results
Holistic multimodel domain analysis revealed that various domains provide unique
information for accurately classifying adolescent suicide attempts. Key domains include negative thoughts and affects about self, interpersonal problems with both parents and peers, victimization and unresolved traumatic experiences, variety of active and passive risk-taking behavior, orientation toward the future, traumatic experiences, lack of resilience and underdeveloped autonomy, and neglect of physiological needs and physical well-being. These indicators were compatible with the six general orientations of the Theory of Adolescent Self-Destruction and its claim that adolescent mental health (well-being and ill-being) requires holistic domain assessment, including physiological, physical, personal, interpersonal, and societal domains as well as orientation about the future. Excluding any of these domains will result in reduction in accuracy of suicide attempt risk estimation. Naturally, it also results in a poorer understanding of adolescent mental health and suboptimal preventive programs to tackle adolescent suicidal behavior.
Importantly, these orientations enabled accurate risk assessment of risk of suicidal
tendencies without relying on suicide-related information, which consequently, enabled accurate differentiation of suicidal and nonsuicidal self-harm. The results also indicated that adolescents who had not previously attempted suicide but were flagged as high risk faced up to a sevenfold greater risk of attempting suicide within the following two years, implying that these domains could predict the future onset of suicidal behavior. Additionally, those estimated to be at lower risk were more likely to report their lifetime suicide attempts inconsistently, further supporting the theory’s standpoint about the future persistence of suicidal tendencies for this age group. These findings align with the theory's assertion that adolescent suicidal behavior is not inherently rational and does not necessarily follow the rational ideation-to-action progression that has been the dominant framework for understanding the development of suicidal behavior. In this context, the theory views adolescent self-destruction from a dimensional perspective and offers an alternative explanation for adolescent suicidality, accounting for its connection to other lethal forms of selfdestructive
behavior that precede or co-occur with suicidal behavior, yet, without a conscious desire or intent to die.
Conclusion
The research presented in this thesis underscores that the development, onset, and
persistence of adolescent suicidal tendencies are predictable through holistic multimodal domain analysis, without relying on any suicide-related information. This holistic risk estimation approach can predict onset and persistence of suicidal behavior up to two years in advance and differentiate highly comorbid conditions seen in suicidal adolescents. The findings were compatible with the theory of adolescent suicidality and self-destruction and partially compatible with the interpersonal theory, the three-step theory, and the strain theory. These findings suggest that the risk and protective factors identified by the posited theory are significant and should be further examined in other cultures. On this ground, it is noteworthy that one of the included journal articles examined the reproducibility of this theory using a 30-year-old data collected from Norwegian adolescents in 1992, which further signifies the generalizability of this theory to a different sociocultural context. Finally, this theory stands out in accurately differentiating adolescent suicidal and nonsuicidal self-harm without relying on suicide-related information.