Why adolescents bully: A global review of theory, measurement, and causal evidence

1. Introduction

Bullying in adolescence has become a prominent public health and educational concern worldwide (Arseneault, 2018; Hosozawa et al., 2021). Meta-analytic evidence indicates that across studies, approximately 35% of young people report traditional bullying victimization and about 15% report traditional perpetration, and cyberbullying affects roughly 15% (victimization) and 6% (perpetration) (Modecki et al., 2014). Importantly, exposure is concentrated in vulnerable groups rather than evenly distributed across adolescents (Abregú-Crespo et al., 2024; Hosozawa et al., 2021). The consequences span multiple domains and settings: in school, bullying is associated with disengagement, absenteeism, and lower academic attainment; in health, it predicts higher risk for depression, anxiety, self-harm/suicidality, and substance misuse; and across adolescence-to-adulthood it forecasts poorer social functioning and socioeconomic outcomes (Arseneault, 2018; Fry et al., 2018; Holt et al., 2015). These harms are sufficiently large to be detectable at the population level. In Australia, national burden-of-disease modelling attributed 10.82% of the burden of depressive disorders and 7.83% of anxiety disorders to bullying victimization in childhood/adolescence, corresponding to an estimated 30,656 disability-adjusted life years (DALYs) lost (95% uncertainty interval 19,304-42,260) (Jadambaa et al., 2020). Complementing these health-loss estimates, an Australian cost-of-illness study estimated annual societal costs attributable to bullying victimization of AUD $763 million in 2016 (and >8% of annual mental health expenditure), underscoring that bullying generates substantial economic as well as human costs (Jadambaa et al., 2021). Reflecting this policy salience, UNESCO Member States have designated the first Thursday of November as the “International Day against Violence and Bullying at School, including Cyberbullying,” and the UN General Assembly has repeatedly adopted the resolution “Protecting children from bullying,” explicitly calling for coordinated prevention and evidence-informed responses.

Research has mapped bullying’s sequelae more coherently than its antecedents (Arseneault, 2018; Hong & Espelage, 2012; Kljakovic & Hunt, 2016). Evidence from longitudinal cohorts and meta-analyses indicates that bullying involvement predicts educational disruption like lower achievement, absenteeism, and dropout, and a graded increase in later mental health problems, most consistently depression and anxiety, as well as higher risk of substance use and suicidal ideation/attempts (Christina et al., 2021; Fry et al., 2018; Holt et al., 2015). Outcome research has benefited from the fact that many large cohorts and school datasets routinely collect downstream educational and psychiatric measures, enabling replication across settings (Hosozawa et al., 2021). By contrast, the field’s understanding of why bullying starts, persists, and concentrates in particular adolescents and school ecologies remains comparatively fragmented.

Identifying causes is difficult partly because bullying is developmentally time-bounded and socially contingent. Bullying victimization tends to peak in early-to-mid adolescence, often around middle-school age and then decline, leaving a narrow window to observe upstream changes before roles stabilize or adolescents transition into different peer ecologies (Sjögren et al., 2024). Measurement also varies substantially across studies, which can distort both prevalence and risk estimates (Modecki et al., 2014; Sjögren et al., 2024). For example, in one middle-school study more than twice as many students were classified as bullies using peer nominations (11%) than by self-report (5%), illustrating how informant choice changes who is “counted” as a perpetrator (Branson & Cornell, 2009). Relatedly, instruments that use behavior-based descriptions versus those that rely on the term “bullying” can yield systematically different prevalence rates (Modecki et al., 2014; Sjögren et al., 2024). Finally, reverse causality and reciprocal influence are pervasive: meta-analytic evidence from longitudinal studies indicates that peer victimization predicts increases in internalizing problems over time, but baseline internalizing symptoms also predict later victimization, consistent with feedback loops that blur temporal ordering (Christina et al., 2021; Reijntjes et al., 2010). Together, developmental timing, heterogeneous measurement, and bidirectional processes mean that many “risk factors” could be either true drivers or downstream correlates.

Despite these challenges, clarifying bullying’s determinants is essential for prevention. Existing syntheses identify correlates at multiple levels—individual (e.g., conduct problems, impulsivity, low empathy), family (e.g., harsh parenting, conflict), peer (e.g., association with aggressive friends, low support), school (e.g., poor climate, weak supervision), and broader social contexts (e.g., discrimination and norms condoning aggression)—but they often read as inventories (Hong & Espelage, 2012; Kljakovic & Hunt, 2016). More specific evidence also suggests that marginalization and clinically relevant vulnerabilities can be salient pathways to victimization: the meta-analysis of clinically diagnosed youth noted above implies that bullying victimization in this group affects roughly two in five adolescents over a 12‑month period (42.2%) (Abregú-Crespo et al., 2024). Yet even within the prospective literature, consensus on which upstream predictors reliably replicate is limited. A meta-analysis restricted to longitudinal adolescent studies concluded that relatively few predictors have been assessed repeatedly, identifying prior victimization, conduct problems, social problems, and internalizing problems as predictors of later victimization, and conduct problems, social problems, school problems, and age as predictors of later perpetration (Kljakovic & Hunt, 2016). This pattern implies both opportunity as some predictors do replicate and a major evidentiary gap as many hypothesized causes have not been rigorously tested.

Clarifying the causes of adolescent bullying has practical significance for research, policy, and society at large. From a scientific perspective, a sharper causal framework will help focus future research on the mechanisms that truly drive bullying behavior, enabling more effective use of study designs, e.g. longitudinal, genetic, and quasi-experimental methods, to test those mechanisms (Schoeler et al., 2018; Veldkamp et al., 2019). For policymakers and educators, identifying modifiable upstream causes of bullying is key to designing interventions with maximal impact (Bradshaw, 2015). Rather than relying on reactive or punitive measures, stakeholders can invest in evidence-based prevention targeting the root etiological factors. Indeed, meta-analyses suggest that comprehensive interventions, those that improve overall school ethos while also providing targeted support for at-risk youth, yield the largest reductions in bullying incidence (Gaffney et al., 2021). Conversely, understanding causation can debunk ineffective approaches: evidence that certain risk factors lie beyond individual students shifts the onus away from blaming victims and toward instituting structural changes (Salmivalli, 2010). In the arena of public discourse, then, a clear message can emerge that bullying is not an inevitable “kids will be kids” phenomenon, but a preventable problem shaped by identifiable risk conditions (Fischer & Bilz, 2019). Highlighting modifiable factors reinforces that society can act on bullying through upstream prevention, for example, by addressing economic disparities, cultural tolerance of aggression, or lack of teacher training. In sum, this review’s broader significance lies in informing a more strategic approach to bullying. By mapping out causes and weighing their relative importance, we can guide research funding to critical gaps, advise lawmakers on where interventions will be most effective, and elevate public understanding of bullying as a problem with systemic solutions.

Ultimately, tackling the causal roots of bullying promises not only to reduce peer violence in schools but also to yield long-term gains in youth mental health, educational attainment, and quality of life (Fraguas et al., 2021; Wolke et al., 2013). As Arseneault (2018) observes, effectively addressing bullying in childhood could alleviate a portion of the burden of psychiatric and socioeconomic difficulties in later life, thereby reducing substantial costs to individuals and society. In pursuit of these benefits, a rigorous global examination of bullying’s causes is both timely and necessary. In confronting bullying, therefore, we are not just solving a schoolyard problem; we are investing in the well-being and potential of the next generation. Each step we take toward understanding causes is a step toward a future in which every adolescent can feel safe, valued, and supported in their formative years.

Despite a rapidly growing literature on bullying correlates and intervention effects, the field still faces a practical bottleneck: many widely cited “risk factors” have been supported mainly by cross-sectional associations, measured inconsistently across contexts, and interpreted without a transparent accounting of causal leverage. As a result, it is often unclear which determinants are likely to be true upstream causes—factors that, if changed, would plausibly shift bullying perpetration and/or victimization—and which are better understood as downstream correlates, markers of broader disadvantage, or components of reciprocal feedback loops. This challenge is especially acute in a global perspective, where differences in language, norms, reporting climates, and measurement instruments can create apparent cross-national variation that does not necessarily reflect real etiological differences.

To address this gap, this review synthesizes the determinants of adolescent bullying with an explicit focus on theory, measurement, and causal evidence. Our overarching aim is to move the field from an inventory of correlates to a set of credible causal claims and clearly bounded uncertainties, while also clarifying which propositions are likely to generalize across contexts and which are inherently context-dependent. The intended beneficiaries include (i) researchers seeking sharper causal tests and a coherent theoretical map, (ii) educators and prevention practitioners who need evidence-informed levers rather than diffuse “risk lists,” and (iii) policymakers who must decide where to invest in upstream prevention under real-world constraints.

The novel contributions of this review can be summarized as follows:

1. Conceptual and measurement clarification. We “unpack” what counts as bullying by specifying core definitional criteria, distinguishing major phenotypes (e.g., physical, relational, cyber, identity-based forms), and emphasizing role fluidity and informant discrepancies—because causal inference is only as credible as the construct being measured.

2. A hierarchy-of-evidence lens for bullying etiology. We introduce a transparent tiering of research designs—from cross-sectional association mapping to longitudinal panels, within-unit designs, quasi-experiments, and randomized trials—to distinguish correlates from determinants with stronger causal support.

3. Theory-by-theory causal appraisal. We treat major etiological frameworks (from individual-propensity accounts to peer, school, neighborhood, and cultural theories) as falsifiable causal claims, and we evaluate what the existing empirical literature actually tests for each theory, where evidence converges, where it conflicts, and where causal conclusions remain premature.

4. A global empirical map that links prevalence patterns to inference limits. We summarize cross-national prevalence and form distributions while highlighting how measurement choices and reporting climates shape observed patterns, and we connect these descriptive regularities to theoretical expectations about what may travel across contexts.

5. A forward agenda for globally credible causal knowledge. We identify key gaps—including “cognitive inequality” in where higher-tier evidence is produced—and outline priorities for multi-wave longitudinal designs, quasi-experimental opportunities, and mechanism-focused interventions that can adjudicate directionality and reduce over-reliance on cross-sectional correlations.

The remainder of this review is structured as follows: Section 2 defines bullying and discusses conceptual boundaries, phenotypes, role fluidity, and cross-context comparability. Section 3 reviews major theories of bullying etiology and evaluates their empirical support using an evidence-tiered causal framework. Section 4 presents a global empirical map of bullying prevalence patterns and form distributions and links these patterns to theoretical generalizability and context dependence. Section 5 integrates theoretical and empirical insights into an overall synthesis of what can be credibly claimed about determinants. Section 6 concludes with a forward-looking agenda, emphasizing study designs and methodological pitfalls that must be addressed to build a genuinely global and causally informative science of bullying.

2. What Counts as Bullying: Conceptual Boundaries, Phenotypes, Role Fluidity, and Comparability

Bullying is typically defined as a subset of peer aggression in which one or more perpetrators intentionally harm a target in a relationship marked by power imbalance, with the behavior occurring repeatedly or involving a sustained threat of repetition (Jia & Mikami, 2018; Smith, 2016; Volk et al., 2014). In Olweus’s classic formulation, a student is bullied when they are “exposed, repeatedly and over time, to negative actions” and the situation involves an asymmetric power relationship that makes it difficult for the target to defend themselves (Olweus, 1993). These three criteria, including intentionality, persistence, and power asymmetry, are widely used as the conceptual hallmarks of bullying (Cuadrado-Gordillo, 2011; Vivolo-Kantor et al., 2014), but each introduces boundary cases that matter for research and for practice (Thomas et al., 2015; Vaillancourt et al., 2008).

Bullying has been conceptualized as a heterogeneous set of phenotypes (Juvonen & Graham, 2014; Turunen et al., 2024). Building on Olweus’s foundational work (Solberg & Olweus, 2003), scholars typically distinguish physical, verbal, relational, and, more recently, cyberbullying as digital platforms extend and reshape peer aggression (Raskauskas & Stoltz, 2007; Wang et al., 2009); a further category, bias-based bullying, targets youths because of group membership, e.g., race/ethnicity, religion, disability, gender identity, or sexual orientation (Galán et al., 2021; Lessard et al., 2020; Rider et al., 2023). Because this review aims to evaluate determinants and causal claims, definitional clarity is not merely semantic: unclear boundaries can change who is counted, what mechanisms are inferred, and whether findings generalize (Chang, 2021; Sawyer et al., 2008).

2.1 Core criteria and boundary cases

Power imbalance is the most diagnostic feature separating bullying from other peer aggression (Nelson et al., 2019). “Power” extends beyond physical strength to social status, numerical advantage (many-on-one), access to resources or information, institutional hierarchy (e.g., older vs. younger students), and discursive authority (e.g., who gets to frame harm as “just a joke”) (Bjärehed et al., 2021; Rawlings, 2019a). In digital contexts, power can be amplified by platform affordances: follower counts, administrator privileges, anonymity, and the ability to spread content widely or keep it accessible over time (Macaulay et al., 2022). The practical implication is that power should be assessed relationally, whether the target can realistically resist, exit, or make the behavior stop, rather than inferred solely from visible traits.

These criteria help distinguish bullying from three neighboring categories: conflict, joking/teasing, and one-off aggression.

First, bullying is not synonymous with peer conflict. Conflicts can be hostile and even violent, but they typically occur between parties of roughly equal power who can each defend themselves, negotiate, or disengage (Rodkin & Berger, 2008; Smith, 2016). Bullying, by contrast, is dominative rather than reciprocal: one side unilaterally imposes harm, coercion, or humiliation in a way the other side struggles to stop (Pellegrini & Long, 2002; Vaillancourt et al., 2003). Empirical work suggests that young people themselves rely heavily on cues about power disparity and persistence when labeling a scenario as bullying rather than “a fight” (Nocentini et al., 2010; Vaillancourt et al., 2008). Thus, measurement instruments and school investigations should not merely ask whether aggression occurred, but whether the target was effectively trapped in an unequal relationship.

Second, bullying can be masked as “just joking,” especially in peer cultures where banter, nicknames, or “roasting” are normative (Land, 2003; Rawlings, 2019b; Steer et al., 2020). The surface form of teasing can resemble bullying, so the key distinction is consent and agency. Playful teasing is typically mutual and reversible: participants can opt out, object, or set limits without fear of escalation (Campos et al., 2007; Kowalski, 2000; Kruger et al., 2006). Teasing becomes bullying when it is one-sided, humiliating, or persistent—and crucially, when the target cannot safely make it stop (Campos et al., 2007; Cuadrado-Gordillo, 2011). Asking “Was it meant as a joke?” is often less informative than asking whether the target could say “stop” and whether that request would be respected (Hymel & Swearer, 2015; Thomas et al., 2015). This matters because perpetrators may use “joking” as a rhetorical shield that reinforces power by invalidating the victim’s reaction.

Third, repetition is a central criterion in many definitions, yet “repetition” is not always best interpreted as repeated identical acts by the same perpetrator. Bullying often entails a pattern of sustained harm, which can be produced by repeated attacks, by group participation (different peers taking turns), or by structural persistence (e.g., reputational damage that continues even when the original act was singular) (Pouwels et al., 2018; Salmivalli, 2010). A single incident can therefore function like bullying if it generates an enduring threat or ongoing exposure. For example, an embarrassing image shared once online can be repeatedly viewed, forwarded, and weaponized by others long after the initial post, creating sustained victimization despite a “one-time” origin (Slonje & Smith, 2008; Sticca & Perren, 2013). These cases motivate a flexible interpretation of persistence: for etiological questions and harm assessment, what often matters is the continuity of vulnerability and coercion, not whether the initial act occurred multiple times in person (Olweus, 2013; Volk et al., 2014). The goal is to avoid definitional rigidity that excludes severe, bullying-like experiences simply because repetition is difficult to document.

2.2 Cultural variation and cross-national comparability

Cross-national research on bullying confronts not only linguistic variation but also an epistemic hierarchy in which concepts and instruments developed in the Global North are frequently treated as universal defaults. From the perspectives of Asian criminology and Southern criminology, this “default universalism” is not a neutral move: it reflects unequal power in the production of criminological knowledge and can reproduce cognitive injustice by positioning Asia and the Global South as mere testing grounds for Northern models rather than as sites of concept formation and theory-building (Carrington et al., 2018; Carrington et al., 2016; Liu, 2017, 2021). In the language of bullying research itself, imposing a dominant group’s definitions without interrogating fit risks mimicking the very dynamic under study, an asymmetrical imposition that sidelines less powerful voices and local vocabularies.

Although the intent-power-persistence definition is widely used in research, what counts as “bullying” can drift across languages and cultural contexts (Maunder & Crafter, 2018; Smith et al., 2002). Some languages lack a single direct equivalent term, and near-equivalents may emphasize different behaviors. Cross-cultural work shows systematic differences in which behaviors parents and students include within the concept: some contexts foreground direct physical or verbal harassment, while others treat social exclusion and group-based psychological abuse as the prototypical form (Menesini & Salmivalli, 2017; Smorti et al., 2003). Japanese “ いじめ” (“ijime”), often translated as bullying, is commonly described as group-driven ostracism and psychological aggression, illustrating how local constructs may not map one-to-one onto Anglo-American usage (Akiba, 2004; Kanetsuna et al., 2006). Cultural norms also shape disclosure: in settings where face-saving or group harmony is prioritized, victims may avoid reporting or labeling experiences as bullying to reduce stigma or avoid disrupting the group (Matsunaga, 2010). This means cross-national prevalence differences can reflect definitional and reporting differences as much as true behavioral incidence (Craig et al., 2009; Solberg & Olweus, 2003).

For global research, these issues imply two measurement principles. First, surveys should rely on behaviorally specific items rather than the single label “bullied,” and should provide a standardized definition that highlights power imbalance and persistence (Green et al., 2013; Kaufman et al., 2020). Second, cross-cultural adaptation should be treated as a validity problem, not a translation problem: careful back-translation, cognitive interviewing, and piloting can detect non-equivalent interpretations of items (Behr, 2017; Sperber, 2004). Consistent with Asian and Southern criminology’s emphasis on epistemic inclusion, adaptation should also allow a shared “etic” core module alongside culturally salient “emic” extensions, so that measurement does not pre-define what counts as harm by Northern prototypes alone (Carrington et al., 2016; Liu, 2021). After data collection, measurement invariance testing is essential to judge whether the same latent construct is being measured across groups (Putnick & Bornstein, 2016; Vandenberg & Lance, 2000). When invariance fails, partial invariance or alignment approaches may permit cautious comparison, but researchers should avoid ranking countries or claiming universal effect sizes when item functioning differs materially across contexts (Rutkowski & Svetina, 2013). In short, comparability is an empirical property to be demonstrated, and it is also an epistemic commitment to doing cross-national research without reproducing the global inequalities of knowledge that Asian and Southern criminology have foregrounded.

2.3 Phenotypes of bullying and why classification changes inference

Bullying is best treated as an umbrella category comprising partially overlapping phenotypes rather than a single homogeneous behavior (Kaufman et al., 2020; Volk et al., 2014). Commonly distinguished forms include physical (hitting, pushing, damaging belongings), verbal (insults, threats), relational/social (exclusion, rumor spreading, friendship manipulation), cyber (digital harassment, public shaming, non-consensual sharing), and identity-based bullying targeting marginalized identities (race/ethnicity, religion, gender, sexuality, disability) (Crick & Grotpeter, 1995; Earnshaw et al., 2018; Nansel et al., 2001; Russell et al., 2012; Slonje & Smith, 2008). These forms correlate and co-occur, but they are not interchangeable: they differ in visibility, typical settings, and the social or technological affordances that make them effective (Salmivalli, 2010; Smith et al., 2008).

A parsimonious way to capture these differences is by considering a small set of dimensions. Directness distinguishes overt physical/verbal attacks from indirect relational harm (Archer & Coyne, 2005; Björkqvist et al., 1992). Visibility distinguishes behaviors readily observed by adults from covert or backstage dynamics (Lowry et al., 2016). Mediation distinguishes cyberbullying, which is shaped by platform affordances such as persistence, potential anonymity, and rapid diffusion (Kowalski et al., 2014; Lapidot-Lefler & Barak, 2012). Group embeddedness matters because many relational and identity-based forms rely on peer participation and norm enforcement rather than isolated individual hostility (Salmivalli, 2010; Salmivalli et al., 1996). These dimensions help explain why determinants may differ by form: the traits and contexts that enable covert exclusion are not necessarily the same as those that enable overt physical intimidation, and online harassment can be “scaled” by platform mechanics in ways unavailable offline.

This heterogeneity creates a methodological risk: collapsing forms into a single “bullying” score can bias causal inference. At minimum, four problems follow. (1) Mechanism averaging: if different forms are driven by different processes, pooling produces an average effect that may fit none of them well, or even yields null findings when effects cancel (Ostrov et al., 2019). (2) Confounder mismatch: the relevant confounders (and therefore the required controls and identification strategy) can differ across forms, so a single model is more likely to be misspecified (Bradshaw et al., 2015). (3) Misclassification: respondents may over-include episodic conflict or under-recognize relational exclusion, adding noise that attenuates associations and distorts subgroup comparisons (Modecki et al., 2014; Vaillancourt et al., 2008). (4) Intervention extrapolation: evidence that a program reduces one form may not generalize to others, and policy constraints on visible violence can sometimes displace aggression into less detectable forms (Cross et al., 2011; Salmivalli, 2010). For this review, the implication is conservative: causal claims should be stated at the level at which they are identified—by form, by role (perpetration vs. victimization), and by context—unless evidence supports invariance across forms.

2.4 Bullying roles, role fluidity, and implications for measurement

Bullying is not only a behavior but a social process involving multiple participant roles (Salmivalli et al., 1996; Swearer & Hymel, 2015). In addition to perpetrators and targets, research distinguishes bully-victims and bystanders who may assist, reinforce, remain passive, or defend (O'Connell et al., 1999; Pouwels et al., 2018). The presence of an audience is central: bullying often functions as a public performance through which perpetrators pursue status, dominance, or peer approval (Dijkstra et al., 2008; Olthof et al., 2011). Bystanders’ reactions therefore help constitute the “power imbalance” in practice: reinforcement or silence can reward the perpetrator and isolate the target, whereas credible defending can disrupt the episode and shift classroom norms (Lynn Hawkins et al., 2001; Salmivalli et al., 2011). Network research further indicates that bullying is patterned by peer ecology: social centrality can protect some perpetrators from sanction, while victims often have fewer allies and occupy more peripheral positions (Hodges et al., 1999; Huitsing & Veenstra, 2012; Huitsing et al., 2012).

Roles are also fluid rather than fixed types. Victims in one setting can bully in another, and longitudinal evidence identifies transitions between roles as peer contexts shift (Hymel & Swearer, 2015; Zych et al., 2020). Bully-victims illustrate this most clearly: they are both aggressors and targets, often experiencing elevated psychosocial risk and instability in peer relations (Haynie et al., 2001; Perren & Alsaker, 2006; Toblin et al., 2005). This role fluidity cautions against interpreting cross-sectional labels (“a bully,” “a victim”) as stable traits; doing so can misattribute to personality what may be contingent on school climate, peer networks, and changing opportunities (Faris & Ennett, 2012; Faris & Felmlee, 2011). Conceptually, it aligns with social-ecological frameworks in which bullying emerges from interactions across levels, individual, peer, school, family, and broader culture, rather than being reducible to “bad kids” or “weak kids” (Hong & Espelage, 2012).

Because bullying is relational and context-dependent, measurement choices matter, especially for cross-study and cross-national synthesis (Hymel & Swearer, 2015; Vivolo-Kantor et al., 2014). Informant differences are consequential: self-reports capture private experiences including covert or online victimization but can be shaped by shame, fear, or definitional misunderstandings; perpetration self-reports are especially vulnerable to social desirability (Modecki et al., 2014; Solberg & Olweus, 2003). Peer nominations can identify reputational bullies and silent victims but may miss hidden victimization and raise ethical risks if confidentiality is not protected (Bouman et al., 2012; Branson & Cornell, 2009). Teacher reports tend to underestimate overall prevalence and skew toward observable physical incidents, missing covert relational and online forms (Bradshaw et al., 2007; Craig et al., 2000). Agreement across informants is typically only moderate, implying that bullying is not a single directly observed “fact” but a construct that is partially perspective-dependent (Holt et al., 2007; Ladd & Kochenderfer-Ladd, 2002). Where feasible, multi-informant designs and latent-variable approaches can triangulate more robust estimates (Marsh et al., 2011; Pouwels et al., 2016).

Comparability also depends on reporting climates: schools and cultures differ in whether students feel safe disclosing bullying, whether adults are trusted, and whether reporting is framed as responsible or stigmatized (Blomqvist et al., 2020; Boulton et al., 2017; Helka et al., 2024; Mazzer & Rickwood, 2015; Park et al., 2023). Awareness campaigns or policy shifts can increase reporting without increasing true incidence, complicating trend interpretation (Novin & Rieffe, 2015; Swearer et al., 2010). For these reasons, cross-context synthesis should treat observed prevalence as jointly determined by behavior, norms, and measurement (Chudal et al., 2022; Jetelina et al., 2019).

Taken together, bullying must be “unpacked” before etiological claims can be evaluated. A usable definition emphasizes intentional harm embedded in a power imbalance and sustained vulnerability, while acknowledging boundary cases that hinge on consent, agency, and persistence (Kaufman et al., 2020). Bullying also comprises distinct phenotypes, and aggregating them can obscure mechanisms and distort causal inference (Modecki et al., 2014). Finally, bullying is a relational process with fluid roles, making informant choice, reporting norms, and measurement equivalence central methodological concerns. These conceptual and measurement clarifications set the stage for the next section: evaluating theories of bullying etiology with explicit attention to what the evidence can and cannot support.

3. Theories of bullying etiology and what the evidence actually tests

3.1 Clarifying the causal evidence commitment in the post-GPT era and evidence hierarchy overview

In an era when AI can readily compile lists of bullying risk factors, the value of a review lies in examining what truly drives bullying (Flanagin et al., 2023). Researchers have identified individual, family, peer, and school factors associated with bullying (Clayton R Cook et al., 2010; Hensums et al., 2023; Ttofi & Farrington, 2012). Correlation is not causation, a list of risk factors does not explain why bullying occurs (Benjamin et al., 2018; McShane et al., 2024; Zhao et al., 2010). In the post-GPT landscape, merely listing correlates is trivial; the real challenge is to distinguish descriptive associations from true causal influences on bullying (Editorials, 2023; D. M. Li et al., 2025; Liu & Li, 2024; Thorp, 2023; Zhao et al., 2025; X. Zhao et al., 2024; Y. J. Zhao et al., 2024).

Accordingly, we approach theories of bullying etiology as falsifiable causal propositions. Criminological and developmental perspectives from classic theories posit candidate causes of bullying (Agnew & White, 1992; Akers, 1985; Cohen & Felson, 1979; V. J. Felitti et al., 1998; Hirschi & Stark, 2014; Liu, 2009; Markowitz et al., 2001; Sampson et al., 1997; Wikström & Treiber, 2009). For each theory, we pose a counterfactual question: if one could intervene to change the theory’s key factor, would bullying perpetration and/or victimization rates change as a result (Höfler, 2005; Rubin, 1974; Ttofi & Farrington, 2011)? This “what-if” approach reframes each theory as a testable causal claim; to evaluate each claim, we employ a clear hierarchy of research designs ordered by their strength of causal inference.

We classify evidence into five tiers of increasing causal leverage.

Tier 1 consists of cross-sectional studies, which map associations at a single time point but cannot establish temporal order and are prone to simultaneity or reverse causality (Barnett et al., 2023; Savitz & Wellenius, 2023; Sedgwick, 2014).

Tier 2 comprises prospective longitudinal studies and panels, which establish temporal ordering by tracking individuals over time, but remain vulnerable to time-varying confounding, selection dynamics (e.g., changing peer networks), and residual reverse causality (Daniel et al., 2013; Grimes & Schulz, 2002; Leszczensky & Wolbring, 2019).

Tier 3 encompasses within-unit designs, such as individual fixed-effects models, sibling or twin comparisons, and school fixed-effects, that control for stable unobserved characteristics; these reduce bias from unmeasured baseline differences and have shown that some putative risk factors lose significance once invariant confounds are held constant (Gunasekara et al., 2014; Imai & Kim, 2019; Sommet & Lipps, 2025).

Tier 4 includes quasi-experimental and natural experiment studies (e.g., differences-in-differences, regression discontinuities, policy shocks, or other exogenous changes) that approximate counterfactual contrasts under explicit assumptions, offering stronger causal evidence when those assumptions hold (Bertrand et al., 2004; Imbens & Lemieux, 2008; Imbens & Wooldridge, 2009).

Tier 5 comprises randomized controlled trials and other experiments (such as cluster-RCTs, stepped-wedge trials, or experimental encouragement designs) that actively manipulate hypothesized determinants or mechanisms, providing the most robust tests of causality (Campbell et al., 2012; Hemming et al., 2015; Schulz et al., 2010).

Before proceeding, two clarifications about this evidentiary framework are warranted. First, categorizing studies by tier is not a value judgment on research quality but a pragmatic assessment of causal leverage (Angrist & Pischke, 2010; Hernán, 2018). Even a perfectly executed Tier 1 study cannot answer a Tier 3 question about within-person change, just as a Tier 5 experiment may face limits in generalizability or mechanism identification (Kraemer et al., 1997; Liu & Li, 2024; MacKinnon et al., 2007; Zhao et al., 2022; Zhao et al., 2010). Second, bullying is fundamentally relational and interdependent in nature: changing one student’s behavior can alter peers’ exposure and responses, complicating causal attribution (Dishion & Tipsord, 2011; Rambaran et al., 2020; Salmivalli, 2010). This interdependence means that even gold-standard interventions might produce combined effects without isolating which specific mechanism drives the outcome (Craig et al., 2008; Han et al., 2023; Imai et al., 2013; Jiang et al., 2021; Skivington et al., 2021). With these caveats in mind, the next section maps out the evidence for each theoretical proposition, identifying (i) which proposed causes of bullying have any credible support, (ii) where the key inferential bottlenecks lie, and (iii) which research designs could most effectively advance causal understanding in this field.

3.2 Theoretical evidence hierarchical analysis

3.2.1 Social Learning Theory

Social Learning Theory posits that aggressive behaviors like bullying are learned through observation and reinforcement in social contexts (Bandura, 1977; O'Connell et al., 1999; Salmivalli et al., 2011). Originally formulated by Bandura and others in the context of modeling aggression, this theory holds that children and adolescents can acquire bullying behaviors by imitating role models, especially those who are rewarded for aggression, and through direct reinforcement of their own aggressive acts (Bandura, 1977; Bandura et al., 1961; Foulkes & Foulkes, 1965). In the bullying context, the imitation/reinforcement proposition is that exposure to aggressive models, e.g. bullying peers or older siblings, and receiving encouragement or rewards for aggressive behavior in salient relationships like peers, siblings, caregivers, increase the likelihood of one’s own bullying perpetration (Akers et al., 1998; Grant et al., 2019; Wolke et al., 2015). A related norm propagation proposition holds that group norms which reward dominance and aggression can spread bullying through peer networks: when peer groups or school classes develop norms that endorse or reward bullying, those attitudes transmit via social contagion, leading to higher overall bullying (Dijkstra et al., 2008; Kärnä et al., 2011; Paluck et al., 2016; Rambaran et al., 2020). By the same token, if the social rewards and reinforcement for bullying are removed or counteracted, bullying behavior should diminish (Perkins et al., 2011; Waasdorp et al., 2012). Finally, an important selection versus influence proposition can serve as a falsifiable “wedge” for causal inference: if social learning truly causes bullying, then exogenous changes in peer group exposure should predict subsequent changes in bullying behavior. In other words, holding a student’s own propensity constant, a random or externally driven shift in their peer context ought to result in changes in their bullying involvement later (Akers et al., 1998; Lodder et al., 2016; Shin, 2019). This proposition differentiates social learning from the alternative scenario where observed peer correlations are entirely due to selection or common underlying traits (McPherson et al., 2001; Snijders et al., 2010).

At the first evidence level, peer-group context shows robust associations with bullying-related aggression, but the interpretation hinges on design (Rodkin et al., 2015). In Child Development, Espelage et al. (2003) examined sixth-eighth graders using peer groups identified via social network analysis and estimated multilevel models; they reported substantial within-group similarity in self-reported bullying and fighting, and found that peer-group bullying/fighting was associated with individual behavior even after accounting for baseline levels, consistent with a contextual effect, but also consistent with homophily (students affiliating with similar peers) (Hanish et al., 2005; Lodder et al., 2016; Merrin et al., 2018).

Complementing this, Dijkstra et al. (2008) in the Journal of Abnormal Child Psychology used multilevel regression in a large Dutch cohort (TRAILS; N≈3,300) and showed that when popular adolescents were involved in bullying, the typical “social costs” of bullying were muted: the negative association between bullying and peer acceptance and the positive association with peer rejection were weakened (Cillessen & Mayeux, 2004; Pouwels et al., 2018). This pattern is especially informative for the reinforcement mechanism in social learning theory, because it indicates that bullying can carry greater social rewards or reduced sanctions in contexts where high-status peers set a permissive “popularity norm,” making the behavior more likely to be maintained and socially transmitted (Berger & Caravita, 2016; Garandeau et al., 2019; Salmivalli et al., 2011).

However, stronger designs underscore that peer similarity is not automatically peer causation, and that the social learning claim is often conditional (Aral et al., 2009; Dishion & Tipsord, 2011; Laursen & Veenstra, 2021). In Development and Psychopathology, Huitsing et al. (2014) used longitudinal stochastic actor-based social network models across three elementary schools to study the coevolution of bullying/victimization ties and defending relations over three waves, illustrating how bullying involvement is embedded in changing alliance structures within networks. Even more directly, Sijtsema et al. (2014) in Developmental Psychology tested whether friends become similar in bullying because of influence while explicitly controlling for friendship selection processes using RSiena. They found clear evidence of selection, youth choosing friends similar in bullying, whereas evidence for friends’ influence on becoming more similar in bullying over time was limited to early adolescence and was not consistently supported in late childhood; some influence patterns were described as marginal and moderated by factors such as moral disengagement (Bussey et al., 2024; Killer et al., 2019). Taken together, these studies provide a concrete example of why cross-sectional and even simple panel associations can overstate a general peer-influence interpretation: what looks like “learning from peers” can reflect selection, developmental stage, and other shared determinants.

The clearest support for social learning mechanisms appears at higher evidence levels that manipulate peer norms or reinforcement contingencies. In a randomized field experiment published in Proceedings of the National Academy of Sciences, Paluck et al. (2016) randomized an anticonflict intervention across 56 schools (over 24,000 students) by training a small “seed” group of students to take a public stance against conflict; compared with control schools, disciplinary reports of peer conflict were reduced by about 25% over one year, with stronger effects when seeds were socially salient “referent” students. Similarly, Kärnä et al. (2011) reported in Child Development a large cluster-randomized evaluation of the KiVa program (78 schools; Grades 4-6) and found consistent beneficial effects after nine months on multiple outcomes including reductions in self- and peer-reported victimization and self-reported bullying. These intervention findings are especially valuable for a social learning account because they operationalize the mechanism as a modifiable lever: changing how peers respond to bullying, e.g., reducing bystander reinforcement and altering perceived norms, can causally reduce bullying-related outcomes (Paluck, 2009; Troop-Gordon et al., 2019; Zambuto et al., 2020).

In conclusion, the empirical record supports social learning as a consequential causal pathway for bullying, but the support is asymmetric across its components. Tier 1-2 literature consistently maps that bullying clusters within peer groups and is patterned by social rewards and status climates, yet such clustering is not automatically peer influence. By contrast, the strongest causal support for social learning comes from Tier 5 evidence: cluster-randomized and field-experimental interventions that shift peer norms and bystander reinforcement, produce measurable reductions in bullying-related outcomes at the school level. Overall, we can state with confidence that peer-ecological incentives and normative reinforcement are modifiable determinants that causally shape bullying prevalence, while the claim that “exposure to bullying peers causes individuals to bully via influence” remains credible but not uniformly established without designs that generate exogenous variation in peer exposure and cleanly separate influence from selection.

3.2.2 Self-Regulation/Self-Control Theory

Self-regulation theory, closely related to Gottfredson and Hirschi’s self-control theory in criminology, posits that individuals with deficits in regulatory abilities are more prone to engage in antisocial or impulsive behaviors, including bullying (Gottfredson & Hirschi, 1990; Muraven & Baumeister, 2000). Key facets of self-regulation include executive functioning, impulse control, and effortful control, in general, the capacity to restrain one’s aggressive impulses and consider long-term consequences. The trait propensity proposition of this theory is that stable deficiencies in self-regulatory capacity, whether due to temperament, neurodevelopmental factors, or learned habits, increase the likelihood of bullying perpetration (Cho & Lee, 2018; Finkel et al., 2009; Unnever & Cornell, 2003; Vazsonyi et al., 2017). A bully, in this view, might be someone who is impulsive, low in empathy, or unable to control anger, thus more likely to lash out repeatedly at weaker peers (Frick & White, 2008; Jolliffe & Farrington, 2011). A secondary proposition can be termed the risk exposure hypothesis: youths with poor self-regulation not only may initiate more aggression, they may also inadvertently increase their own risk of victimization. This is because impulsivity and poor decision-making can lead them into conflictual situations, provoke stronger peers, or generally reduce their vigilance and self-protection, thereby heightening the chance of being bullied by others (Meldrum et al., 2012; H. Zhao et al., 2024). A change hypothesis further implies that if an individual’s self-regulatory ability improves or deteriorates over time, we should observe a corresponding decrease or increase in that individual’s bullying behavior, independent of any stable personal traits. This intra-individual prediction separates the effect of a person’s enduring trait from the effect of fluctuations or development in self-regulation. These propositions collectively underscore that self-regulation is not merely correlationally associated with bullying but is posited as an underlying causal factor that can be tested by looking at both between-person differences and within-person changes (Gottfredson & Hirschi, 1990; Vazsonyi et al., 2017).

Bullying-specific evidence at the lower tiers largely aligns with the meta-analytic baseline but also clarifies boundary conditions (Clayton R. Cook et al., 2010; Kasturiratna et al., 2025; Pratt et al., 2014). For instance, Unnever and Cornell (2003), published in the Journal of Interpersonal Violence, explicitly linked self-control deficits to school bullying and discussed how neurodevelopmental traits such as ADHD may co-travel with regulatory difficulties, sharpening the “stable propensity versus situational fluctuation” boundary problem, that is whether low self-control is a trait-like liability, a state-like depletion, or a composite of both. Cross-national work also supports transportability but warns against assuming invariance: Vazsonyi et al. (2012) in the European Journal of Developmental Psychology tested low self-control in relation to cyberbullying across 25 European countries, treating self-control as a cross-cultural risk factor while simultaneously documenting contextual heterogeneity in effect magnitudes. Studies that explicitly embed regulatory capacity in broader ecological exposure further illustrate why a single-factor reading is risky: Holt et al. (2014) in the Journal of Criminal Justice modeled self-control alongside neighborhood disorder when predicting bullying victimization, reinforcing the idea that individual regulatory capacity and situational exposure jointly shape risk. Likewise, Chui and Chan (2013) in Child Abuse & Neglect examined both bullying perpetration and victimization among Macanese adolescents, supporting a “dual-path” formulation in which low self-control can elevate perpetration risk while also increasing victimization risk through heightened exposure and poorer conflict navigation.

At the same time, the same high-quality literature that supports a link between low self-control/regulatory difficulty and bullying also repeatedly signals why the association should not be narrated as a simple one-way causal arrow (Moon & Alarid, 2014; Vrijen et al., 2024). First, reverse pathways are plausible and often empirically competitive: sustained involvement in bullying can plausibly erode regulatory capacity via chronic stress, conflict escalation, school disengagement, and reinforcement of impulsive scripts, processes that would make “low self-control” partially an outcome of bullying trajectories rather than a pure antecedent (Herd et al., 2024; Ouellet-Morin et al., 2011). Second, time-varying contextual covariation is a structural threat to inference: family disruption, deviant peer affiliation, classroom climate, and neighborhood disorder can simultaneously move both self-regulation and bullying involvement, so Tier 1-2 associations can overstate a direct effect unless reciprocal dynamics and shared upstream drivers are explicitly modeled (Bowes et al., 2009; Holt et al., 2014). Longitudinal trajectory work makes this caution concrete: Cho et al. (2019) in Children and Youth Services Review used a latent growth curve framework to move beyond static “difference” claims and toward change processes, while Cho and Lee (2018) in the Journal of Interpersonal Violence jointly modeled low self-control and delinquent peer associations, implicitly underscoring that self-control theory and social-learning mechanisms are often nested rather than mutually exclusive. A careful etiological narrative, therefore, should present self-regulation as a consistent risk marker with plausible causal influence, while also being explicit that without designs addressing reciprocal causation and shared contextual determinants, the literature can too easily convert correlation into an unearned one-directional causal story (Arseneault, 2018; Lucas, 2023).

Notably, even though the evidence at a higher level is still not yet abundant enough, evidence for this mechanism also reaches the upper tiers of the hierarchy: quasi-experimental evaluations of programs that explicitly train adolescents’ self-management and emotion regulation among identified perpetrators report subsequent reductions in school bullying behavior (Song & Kim, 2022). At the experimental tier, a cluster randomized mindfulness intervention produced gains in self-control alongside declines in bullying, with self-control operating as a key mediator of the intervention effect (Liu et al., 2022).

Overall, the evidence supports self-regulation deficits as a reliable antecedent of bullying perpetration and a plausible intervention target. While reciprocal processes and contextual confounding mean the causal story is not fully closed, the combination of consistent prospective associations and emerging intervention evidence suggests self-control is a modifiable mechanism that can reduce bullying when strengthened. Future work should therefore prioritize designs that isolate within-person change and test mediation, rather than accumulating additional cross-sectional correlations.

3.2.3 Social Bond Theory

Social bond theory, rooted in Hirschi’s control framework, proposes that attachment, commitment, involvement and belief embed adolescents in conventional institutions and raise the costs of violating rules (Chapple et al., 2005; Costello & Laub, 2020). In bullying, the most relevant bonds are emotional connection to parents and school, investment in schooling as a valued trajectory, participation in adult-supervised activities, and endorsement of school norms as legitimate (Bowes et al., 2009). When such bonds are weak, bullying should become less costly and less constrained, and informal social control should be thinner because fewer ties create fewer everyday opportunities for monitoring and accountability (Brookmeyer et al., 2006; Goldweber et al., 2013; Hirschi, 2017; Sampson, 1986). The theory is also falsifiable: support is most diagnostic when earlier bonding predicts later bullying involvement net of prior bullying, whereas consistent evidence for the reverse pathway, bullying involvement preceding declines in bonding, would challenge a simple unidirectional account (Buhs et al., 2006; Hemphill et al., 2012; Li et al., 2020; Popp & Peguero, 2012).

A useful Level-1 evidence for the “bonding as protection” claim comes from population evidence that is often treated as a conceptual base for school/family connectedness as a general protective factor (Bond et al., 2007; Libbey, 2004; Rose et al., 2022). In a cross-sectional analysis published in JAMA, Resnick et al. (1997) analyzed interview data from the U.S. National Longitudinal Study of Adolescent Health (Add Health; n=12,000 adolescents in grades 7-12) and reported that parent-family connectedness and perceived school connectedness were protective across multiple health-risk domains, including violence. While this study does not directly test bullying and is cross-sectional, it is a strong exemplar of the broader empirical pattern that “attachment to family and school” co-travels with lower adolescent harm-precisely the type of pattern social bond theory generalizes to bullying involvement.

At Level 2, bullying-specific longitudinal studies reinforce the idea that bonds are embedded in a multi-layer ecology (family-school-neighborhood), but they also show that “bond” effects are not always cleanly separable from correlated risks (Han et al., 2019; Juvonen & Graham, 2014; Swearer & Hymel, 2015). For example, Bowes et al. (2009), published in the Journal of the American Academy of Child & Adolescent Psychiatry, used a nationally representative UK birth-cohort design (E-Risk Longitudinal Twin Study; teacher and mother reports of bullying involvement in early childhood) to test whether school, neighborhood, and family factors predict bullying involvement over and above children’s early behavior problems. They found that socioenvironmental factors were associated with children’s risk of being a victim, bully, or bully-victim even after accounting for behavioral vulnerabilities. This kind of evidence supports an ecological reading of social bond theory: “weak bonds” are rarely just an individual disposition; they often sit alongside family conflict, neighborhood disorder, and school structural features that jointly condition bullying risk (Hong & Espelage, 2012; Låftman et al., 2017).

Evidence also suggests that bonding-related predictors can vary by bullying form and can weaken under stricter model specifications, which is exactly the type of nuance that helps avoid overstating one-way causality (Camacho et al., 2023; Kowalski et al., 2014; Marciano et al., 2020). In a prospective study in the Journal of Adolescent Health, Hemphill et al. (2012) followed Australian adolescents and examined Grade-7 predictors of later cyber and traditional bullying perpetration. While indicators linked to school disengagement (e.g., low school commitment) can appear predictive in partially adjusted models, some of these “bond” associations attenuate once a fuller set of correlated individual/family/peer risks is included whereas other predictors, like prior relational aggression, family conflict, academic failure, and prior victimization/perpetration for traditional bullying, remain more robust. This pattern is useful to state explicitly because it illustrates why Level-2 evidence can be “promising but uneven”: the apparent protective role of school bonding is sometimes sensitive to whether one treats it as an independent driver versus a correlate of broader co-occurring risks (Kljakovic & Hunt, 2016; Mathur & VanderWeele, 2020).

The evidence offers qualified support for social bond theory in bullying. Attachment to parents and school, commitment to schooling, and perceived legitimacy of school norms are consistently associated with lower bullying involvement, and some prospective studies suggest that earlier connectedness predicts later bullying outcomes. Yet the strongest longitudinal evidence also indicates substantial reciprocity, bullying involvement, particularly victimization, can undermine belonging, trust, and school engagement, making “weak bonds” as plausibly an outcome as a cause. Moreover, bond indicators often attenuate under fuller adjustment, implying they may partly proxy broader family, peer, and school risks rather than operate as independent drivers. The most defensible conclusion is therefore a bidirectional, ecological account: bonding is a meaningful protective resource and a plausible lever for prevention, but it is embedded in—and can be eroded by—the wider risk ecology that sustains bullying.

3.2.4 General Strain Theory

As a theoretical starting point, Robert Agnew’s General Strain Theory (GST) was originally formulated as a testable causal chain in Criminology: exposure to strain, broadly, the experience of aversive events or blocked goals, increases negative affect, especially anger, thereby elevating the likelihood of “coping” through aggression and other deviant behaviors (Agnew, 1992). A key refinement that matters for bullying is that GST does not treat all stressors as equally “criminogenic” (Broidy & Agnew, 1997; Moon et al., 2009). In Journal of Research in Crime and Delinquency, Agnew (2001) specified that strains are most likely to precipitate crime when they are high in magnitude, perceived as unjust, associated with low social control, and create pressure or incentive for maladaptive coping. Translating this to bullying etiology makes the theory more falsifiable and more specific: strains such as peer victimization, chronic peer rejection/humiliation, harsh or conflictual family environments, and salient academic failures are not merely “correlates,” but candidate upstream exposures that should temporally precede bullying perpetration; operate via measurable negative emotions, and show conditional effects depending on coping resources (e.g., social support, self-control) (Agnew, 2001; Clayton R. Cook et al., 2010; Hemphill et al., 2012; Walters & Espelage, 2018).

At Tier 1, cross-sectional tests have commonly treated GST as a mechanism-focused mode, by measuring strains, negative affect, and bullying outcomes within the same survey wave and examining whether the predicted mediation pattern is evident (Agnew, 1992; Agnew, 2001; Agnew & White, 1992; MacKinnon et al., 2007; Maxwell & Cole, 2007). For example, Hay et al. (2010) in Youth Violence and Juvenile Justice explicitly situated traditional bullying and cyberbullying within a GST framework and reported that victimization/negative life events (strain indicators) were associated with deviance and bullying-related outcomes in ways broadly consistent with the “strain → negative affect → problem behavior” logic. Similarly, Patchin and Hinduja (2010) in Youth & Society treated GST as a falsifiable account for both traditional and “nontraditional” bullying, and their results are useful precisely because they show how GST can be only partially supported: while some strains and emotional correlates align with GST expectations, the full mediation/chain is not always uniformly recovered across bullying forms or model specifications.

Level 2 studies that leverage temporal ordering offer a more discriminating appraisal, and they illustrate both GST’s promise and its unresolved points (Agnew, 1992; Aseltine et al., 2000; Moon et al., 2008). In Korean panel data, school-generated strains such as teachers’ punishment and examination stress predict subsequent bullying, but the overall explanatory reach of GST is modest, underscoring that strain is only one piece of bullying’s etiology (Juvonen & Graham, 2014; Lee & Larson, 2000; Moon et al., 2008). Using a larger South Korean longitudinal sample, peer victimization and parent-child conflict prospectively predict bullying, yet the expected conditioning effects and the anger mediation pathway receive limited support (Espelage et al., 2014; Moon et al., 2010; Reijntjes et al., 2011). In a U.S. three-wave design, bullying victimization predicts later maladaptive outcomes partly through negative emotions, aligning with GST’s broader strain-emotion logic even when bullying itself is treated as the strain (Hay & Meldrum, 2010; Holt et al., 2015). More targeted longitudinal mediation tests of the victimization-perpetration pathway suggest that the link may be carried less by anger than by cognitive hostility, implying that GST’s emotional mechanism may require specification rather than unqualified acceptance (Gao et al., 2024; Malamut & Salmivalli, 2021; Walters & Espelage, 2018). Longitudinal evidence that punitive parenting predicts later bullying perpetration through peer processes further supports the relevance of chronic interpersonal strain, while also pointing to peer mechanisms that GST typically treats as contextual (Hong et al., 2017; Moon et al., 2008).

At Tier 3, Park and Metcalfe (2019) provide rare within-individual evidence by using five waves of the Korean Youth Panel Survey to test whether year-to-year changes in bullying victimization are associated with changes in delinquent coping responses, including bullying. Even so, Tier 3 designs cannot fully rule out time-varying confounding (e.g., shifting peer ecologies or school climates that may jointly move victimization and bullying), so stronger counterfactual leverage is still needed. Future tests of GST in bullying should prioritize stronger designs, like quasi-experimental variation in salient strains and well-powered school-based experiments that reduce strain or strengthen coping, paired with repeated measurement to evaluate temporal ordering and mediation.

Overall, the weight of evidence provides moderate support for GST as an account of bullying etiology: a range of salient strains, especially peer victimization, chronic interpersonal conflict, and school-generated stressors, often precede later bullying involvement, and the broader strain-maladaptive coping logic appears replicable across contexts. However, the evidence for GST’s canonical emotional mechanism is uneven, anger mediation is frequently weak or inconsistent, and some longitudinal studies suggest that cognitive hostility and peer processes may be more proximal pathways than anger per se. Moreover, reverse causality remains a recurring concern because bullying involvement can itself function as a strain that erodes adjustment. The most defensible conclusion, therefore, is that GST captures a real adversity-to-bullying pathway, but its mechanism requires tighter specification and stronger designs to distinguish causal effects from reciprocal dynamics.

3.2.5 Collective Efficacy Theory

Collective efficacy can be read as a mechanism-focused successor to classic social disorganization theory. The social disorganization tradition emphasized how structural conditions, concentrated disadvantages, residential instability, weak institutional capacity, undermine informal social control, thereby elevating violence and disorder (Bursik Jr, 1988; Sampson et al., 1997). The collective-efficacy reformulation shifts the explanatory center of gravity toward a measurable social process: social cohesion/trust paired with shared expectations for informal intervention. This conceptual move matters for bullying because it reframes “context” as a causal mechanism, while also cautioning that “more ties” are not automatically protective unless they are coupled to prosocial norms and readiness to regulate harmful behavior(Kubrin & Weitzer, 2003; Sampson et al., 2002).

Collective efficacy theory, originally formulated in the context of neighborhoods, emphasizes the role of shared trust and mutual willingness to intervene for the common good in regulating antisocial behavior (Morenoff et al., 2001; Sampson et al., 1997). In a school setting, collective efficacy refers to the collective capacity of members to maintain a safe, orderly environment through cohesion and informal social control. The informal social control proposition here is that higher collective efficacy in a classroom or school, meaning students and staff have strong social cohesion and are willing to intervene to stop misconduct, will result in lower rates of bullying (Sampson et al., 1997; Sjögren et al., 2020; Williams & Guerra, 2011). Essentially, when a peer group or a school community has a norm of “we look out for each other and won’t tolerate bullying,” potential bullies are deterred by the expectation that others will intervene or disapprove. Conversely, in a class with low collective efficacy, bullies find easy opportunities because bystanders remain passive and there is a lack of unified opposition to aggression (Darley & Latane, 1968; Sjögren et al., 2020; Thornberg et al., 2020). Another proposition can be termed contextual effects beyond composition: collective efficacy theory predicts that differences in bullying rates between classrooms or schools are not just due to which individuals are present, but also due to emergent properties of the group including climate, norms, shared efficacy (Garner & Raudenbush, 1991; Raudenbush & Sampson, 1999; Williford & Zinn, 2018). In other words, even controlling for individual risk factors, some classes may have less bullying because of a strong anti-bullying collective ethos, whereas others have more due to a “culture of silence” or disunity that allows bullying to flourish. Finally, implicit in this theory is a causal ordering: improvements or declines in collective efficacy should precede corresponding changes in bullying levels. If bullying changes first and collective efficacy only shifts as a reaction, it would challenge the direction of causality posited, though bidirectionality is possible. Another test is whether interventions that boost collective efficacy, like building cohesive student teams or empowering bystanders, lead to bullying reductions; failure of such interventions would call into question the theory’s practical import (Kärnä et al., 2011; Ttofi & Farrington, 2011).

The empirical anchor for collective efficacy comes from the neighborhood literature that first operationalized the construct and demonstrated its association with violence in multilevel models (Brunton-Smith et al., 2018; Mazerolle et al., 2009; Raudenbush & Sampson, 1999; Sampson et al., 1999; Weisburd et al., 2020). In the canonical Chicago study published in Science, Sampson et al. (1997) aggregated resident surveys to 343 neighborhoods and showed that collective efficacy was negatively associated with violence, and that the associations of concentrated disadvantage and residential instability with violence were largely mediated by collective efficacy. Morenoff et al. (2001), writing in the American Journal of Sociology, extended this framework by embedding collective efficacy in spatial dynamics: inequality and spatial proximity to violence were consequential for homicide patterns, and low collective efficacy remained a key social process associated with higher homicide, highlighting that “context” operates through both social processes and spatial interdependence rather than being fully contained within administrative borders.

At the same time, the neighborhood evidence that motivates collective efficacy is largely cross-sectional or time-lagged observational, which makes causal direction difficult to establish. Reviews of “neighborhood effects” therefore foreground measurement and identification problems, especially reverse causality and structural confounding, as central challenges for collective-efficacy claims (Garcia et al., 2007; Lanfear et al., 2020; Morenoff et al., 2001; Sampson et al., 1999). In addition, the link between “social organization” and violence need not be monotonic. Browning et al. (2004) in Social Forces articulate a “paradox of social organization”: dense local networks may coexist with high violence when networks also supply social capital for offenders, which can attenuate the expected protective effect of collective efficacy. This line of work supports a more neutral formulation in which cohesion and connection are not presumed uniformly beneficial; what matters is whether networks are mobilized toward prosocial informal control versus tolerance, accommodation, or enablement of aggression.

Longitudinal panel evidence reinforces the need to treat causal direction as an empirical question rather than a presumption. Hipp and Wickes (2017) in the Journal of Quantitative Criminology provide a template for explicitly testing reciprocal dynamics; using repeated neighborhood data, their results do not recover a simple direct collective-efficacy → violent crime relationship over time, instead pointing to reciprocal dynamics involving disadvantage and violence and an indirect relationship between collective efficacy and violence. For bullying research, the parallel is immediate: cross-sectional school “collective efficacy” or climate measures can be consequences of bullying as well as causes of bullying (Forsberg et al., 2024; Waasdorp et al., 2022).

The current best reading is that collective efficacy is better supported as a concurrent marker and potential mediator/maintenance factor of bullying-relevant contexts than as a clearly demonstrated antecedent. The next decisive step is therefore not more cross-sectional associations, but multi-wave longitudinal designs that model feedback and time-varying confounding (e.g., school/class fixed effects, reciprocal cross-lagged/RI-CLPM approaches), ideally supplemented by quasi-experimental or intervention-driven shocks that can test whether changes in collective efficacy reliably precede changes in bullying while holding composition constant.

3.2.6 Labeling Theory

Labeling theory contends that formally designating a young person as a “bully” or a “problem student” can itself foster subsequent bullying through stigmatization and altered self-identity (Link & Phelan, 2001; Major et al., 2005; Thomas & O'Neill, 2021). Originally formulated in sociology to explain deviance, the theory posits that official sanctions attach a negative label that adolescents may internalize, leading to a self-fulfilling prophecy of continued misbehavior (Bernburg et al., 2006; Paternoster & Iovanni, 1989; Rosenhan, 1973). In a school context, being singled out as a bully can undermine one’s attachment to positive peer groups and encourage association with similarly labeled deviant peers, thereby amplifying delinquent behavior (Bernburg et al., 2006; Duxbury & Haynie, 2020). The core causal claim is that the very act of labeling increases the likelihood of future bullying by shaping the student’s self-concept and by eliciting negative expectations from others (Matsunaga, 2010; Mowen & Brent, 2016). An extension of this argument is the displacement hypothesis, which suggests punitive anti-bullying policies might suppress visible, overt bullying while inadvertently shifting the aggression to more covert forms or provoking retaliatory aggression by the labeled student (Borgwald & Theixos, 2013; Hall, 2017). A falsifiable prediction, therefore, is that absent any true difference in baseline behavior, a student who is formally disciplined and labeled would show more subsequent bullying than an otherwise similar student who avoided the label; conversely, if labeling has no causal effect, then changes in school labeling policies should not alter bullying trajectories net of individual risk factors.

At the first evidentiary tier, cross-sectional findings can appear to “confirm” labeling theory because disciplined or negatively regarded students often report higher concurrent involvement in bullying and broader externalizing behavior. Yet these associations are intrinsically compatible with selection and reverse causality: exclusionary discipline is strongly endogenous to prior aggression, and reputational marking can simply be an accurate social reading of persistent conduct (Hirschfield, 2018; Mowen et al., 2020). Indeed, even within bullying-related processes, conflict can precede labeling-like outcomes: recent evidence indicates that peer victimization can predict subsequent suspension, partly via behavioral and emotional responses to being targeted, underscoring that “discipline” may sometimes be an effect of bullying dynamics rather than their cause (Talaugon et al., 2026). Moreover, cross-sectional work on school disciplinary climates complicates a simple “punishment backfires” narrative. When discipline is perceived as structured and supportive, an “authoritative” climate, schools report lower bullying and victimization, a pattern consistent with deterrence or collective norm regulation rather than deviance amplification (Gregory et al., 2012). Critiques of zero-tolerance approaches nonetheless highlight that punitive responses can be stigmatizing and potentially counterproductive, including by shifting bullying into less visible forms (Borgwald & Theixos, 2013). Taken together, first-tier evidence is best read as descriptive of co-occurrence: labels, discipline, and bullying cluster in the same students and settings, but cross-sectional data rarely identify whether labeling is an independent cause, a correlate of chronic risk, or a contingent process that matters primarily through exclusion from prosocial networks (Duxbury & Haynie, 2020; Gerlinger et al., 2021; Hirschfield, 2018; Jacobsen, 2020).

Tier 2 designs provide a more discriminating, though still incomplete, test by establishing temporal ordering and attempting to adjust for baseline risk. In longitudinal delinquency research that is directly informative for school-based labeling claims, first juvenile arrest predicts later self-reported offending and later arrest even in propensity-score matched comparisons; importantly, rearrest effects appear substantially larger and partly independent of changes in self-reported offending, consistent with “secondary sanctioning” (heightened surveillance and harsher societal response) in addition to “secondary deviance” (behavioral change) (Beardslee et al., 2019; Liberman et al., 2014). School-centered longitudinal studies show parallel patterns: cumulative suspensions predict greater likelihood of later arrest even after adjustment for prior behavioral indicators, supporting the possibility that formal sanctions can function as a negative turning point (Mowen & Brent, 2016). Mechanistic longitudinal work further suggests that sanctioned youth can experience interpersonal exclusion and friendship-network disruption, which is plausibly one route through which sanctions become criminogenic rather than corrective (Duxbury & Haynie, 2020; Jacobsen, 2020). However, the relevance of these findings to bullying specifically remains limited, because most studies operationalize labeling via sanctions that are tightly confounded with aggressive behavior and because educational harms of sanctioning can occur even without a specific “bully” identity being attached (Sampson, 1986; Sampson et al., 1999). Stronger tests for bullying will require designs that separate being “named” from being aggressive, for example, quasi-experiments around policy changes in disciplinary labeling, within-person models comparing individuals before and after a labeling event, or interventions that replace stigmatizing responses with restorative practices while holding baseline behavior constant (Acosta et al., 2019; Anyon et al., 2016; Fisher & Widdowson, 2023; Hirschfield, 2018; Hirschi, 2017; Paternoster & Iovanni, 1989). Until such evidence accumulates, labeling theory remains a plausible but underidentified explanation for bullying escalation: its empirical support is suggestive, its counterfactual predictions are clear, and its vulnerability to reverse causality and selection should be treated as a central interpretive constraint (Gregory et al., 2012; Liberman et al., 2014; Mowen & Brent, 2016; Talaugon et al., 2026).

In sum, while observational evidence aligns with labeling theory, more convincing causal tests, like policy experiments, natural experiments, or targeted interventions, are needed. If future quasi-experiments find that reducing or removing deviant labels without changing the individual’s prior behavior leads to lower subsequent bullying, that would bolster the claim that labeling is a true causal influence, not a mere correlate of persistent misbehavior.

3.2.7 Adverse Childhood Experiences (ACEs)

The ACEs framework treats abuse, neglect, and household dysfunction as cumulative early exposures that can alter stress responsivity and socioemotional development, thereby elevating risk for later interpersonal aggression and victimization(V. J. M. D. Felitti et al., 1998; McLaughlin & Sheridan, 2016; McLaughlin et al., 2014; Shonkoff et al., 2012). Translated to bullying, it implies a graded association between ACE accumulation and involvement as perpetrator, victim, or bully-victim, operating partly through downstream dysregulation or psychopathology rather than being a mere correlate of socioeconomic disadvantage (Arseneault, 2018; Merrin et al., 2024; Reuben et al., 2016). The framework is falsifiable: prospectively measured adversities should predict later bullying after accounting for prior behavior, and the association should persist under designs that better control stable family background; failure would suggest ACEs index broader liability rather than causal exposure (Arseneault, 2018; Baldwin et al., 2019; McLaughlin et al., 2014; Merrin et al., 2024).

From a broader social disorganization lens, ACEs are not randomly distributed: structural disadvantages and community disorder can increase families’ exposure to stress, violence, and instability, thereby raising the probability that children accumulate adverse experiences long before they enter the school peer ecology. In this sense, ACEs provide an upstream “cumulative exposure” framework rather than a single-mechanism theory of bullying. The modern empirical tradition begins with Felitti and colleagues’ foundational ACE study in the American Journal of Preventive Medicine (V. J. M. D. Felitti et al., 1998), which operationalized ACEs as a count-based index of childhood abuse and household dysfunction and demonstrated graded (dose-response) associations with a wide range of later-life health and behavioral risks. Transposed to bullying etiology, the central causal proposition is not simply that “children from adverse homes become bullies,” but that cumulative adversity can shape developmental capacities that are directly relevant to peer aggression and vulnerability, including emotion regulation and threat sensitivity, aggressive scripts and normative beliefs about coercion, attachment insecurity, and impaired peer relationship competence, thereby increasing the likelihood of later involvement in bullying as a perpetrator, a victim, or a bully-victim (Hong & Espelage, 2012; Merrin et al., 2024). A key falsifiable implication is the dose-response claim: net of baseline behavioral problems and other confounds, greater cumulative exposure should predict higher subsequent bullying involvement, ideally through measurable mediators (e.g., dysregulation, externalizing pathways, or social skill disruptions), rather than functioning merely as a generic marker of disadvantage.

At Tier 1, the bullying literature shows robust cross-sectional co-variation between ACE exposure and bullying outcomes, but most evidence remains weak on temporal ordering. In large adolescent and population-based surveys, greater cumulative ACE exposure is associated with higher likelihood of bullying involvement across roles. For example, Forster and colleagues found that adolescents reporting more cumulative ACEs showed elevated odds of both school-based victimization and perpetration, consistent with a graded (dose-response) pattern (Forster et al., 2017). Complementary evidence comes from nationally representative caregiver-report data: in the 2020-2021 National Survey of Children’s Health, Everett and colleagues constructed a 10-item cumulative ACE scale and reported higher odds of both bullying victimization and bullying perpetration as ACE counts increased in adjusted logistic model (Dixon Everett et al., 2025). Nonetheless, because ACEs and bullying are typically measured contemporaneously in these surveys and may share reporters, Tier 1 findings remain vulnerable to shared-method bias, reverse temporality, and unmeasured confounding, so they are best interpreted as strong risk mapping and hypothesis support, not definitive causal confirmation.

Beyond Tier 1, there is a small but growing higher-tier literature that more directly tests temporality and alternative explanations. In prospective cohort analyses of the ABCD study, ACE exposure shows a dose-response association with cyberbullying victimization assessed two years later (Nagata et al., 2023), and in the Growing Up in Scotland cohort, prospectively measured ACE indicators were used to predict bullying victimization and bullying perpetration at age 14 (Sapouna, 2025). Moving toward Tier 3, multi-wave panel models that separate stable between-person differences from within-person change report reciprocal associations between emotional maltreatment and bullying perpetration/victimization over time, consistent with dynamic spillover processes (Li et al., 2021). Tier 4 evidence is rarer, but quasi-experimental work leveraging exogenous shocks such as Hurricane Katrina documents changes in relational and overt bullying among affected youth compared with nonaffected peers (Terranova et al., 2009). Finally, Tier 5 randomized family-based interventions that strengthen parenting practices and youth coping skills have demonstrated reductions in bullying involvement, supporting the claim that ACE-linked mechanisms are modifiable determinants rather than fixed correlates (Healy & Sanders, 2014).

The ACEs framework currently has one of the strongest empirical foundations among proposed upstream determinants of bullying. Across large population samples, prospective cohort evidence, and the limited but informative quasi-experimental and intervention findings, cumulative childhood adversity shows a consistent, graded association with later bullying involvement. While residual family-level confounding can never be eliminated entirely in observational work, the convergence of temporality, dose-response patterning, and evidence of modifiability supports treating ACE exposure as a genuine antecedent contributor to bullying risk. Therefore, the field’s priority should shift from asking whether ACEs matter to determining how to prevent ACE exposure and, where prevention is not feasible, how to interrupt the downstream pathways through which adversity becomes peer aggression and vulnerability. This points directly to trauma-informed bullying prevention that is integrated with family- and community-level strategies and with targeted supports that reduce dysregulation and hostile cognitive styles in high-ACE youth.

3.2.8 Relationism Theory

Liu’s relationism theory was developed within his broader “Asian paradigm” in criminology and criminal justice. Rather than treating collectivism as a vague cultural label, Liu argues that the common denominator of many Asian collectivist patterns is the degree to which people are embedded, emotionally and behaviorally, in meaningful personal and group ties. He terms this level of embeddedness “relationism,” and proposes that cross-cultural differences in relationism are a basic source of differences in cultural values, thought patterns, and fundamental concepts of wrongdoing and justice (Liu, 2016). Building on this foundation, Liu specifies three interlocking relationist cultural values: (a) attachment to primary relations and inner groups, (b) honor/face and reputation within relationships, and (c) harmony and conflict avoidance; he links these values to a holistic thinking mode that interprets behavior in relation to the broader relational field rather than as attributes of isolated actors (Liu, 2016; Liu, 2021). In Liu’s synthesis of Asian paradigm development, relationism is framed as a basic paradigm widely shared in Asian societies, with four key elements, attachment to family and community, honor, harmony, and holistic thinking, and it is presented as the newest conceptual innovation emerging from a three-stage roadmap for paradigm building (Liu, 2017, 2021). More recently, Liu extends relationism into a general theory-building program in criminal justice: he critiques the dominant “monotonic cause” paradigm and proposes a more general “comparison paradigm” in which causal processes are constructed and modeled in paired, contrasting terms, most centrally, relationist versus individualist concepts and processes, to capture the mixed nature of social reality across cultural contexts (Liu, 2024; Liu & Li, 2024; Zhang & Liu, 2023). For bullying, this framework is relevant because it pushes explanation away from individual deficits and toward the relational field: bullying becomes intelligible as a relational action that manages belonging and attachment, protects or contests face and reputation, and polices harmony and group order. A relationist reading therefore expects bullying to be produced by mixtures of relationist and individualist motives and to vary with the relational ecology in which actors are embedded, rather than remaining stable across contexts for the same individuals.

Direct applications of Liu’s relationism theory to school bullying research remain limited, and most studies that describe bullying as “relational” do not operationalize Liu’s specific relationist cultural values (attachment, honor/face, harmony, holistic thinking) or test the paired predictions implied by the comparison paradigm. For bullying specifically, the strongest support is therefore indirect and currently best characterized as Tier 1-2: a substantial body of work shows that bullying is structured by peer relationships, bystander roles, and group hierarchies, and that perpetrators’ and targets’ positions in classroom networks matter for who bullies whom and with what social consequences (Faris & Ennett, 2012; Salmivalli et al., 1996; Salmivalli et al., 2011). These findings are consistent with relationism’s core claim that outcomes are not reducible to isolated actors. However, what remains under-aligned with Liu’s theory is the cultural and comparative content: bullying studies rarely test whether face concerns, harmony norms, and attachment to inner groups condition bullying mechanisms, or whether relationist processes can be empirically distinguished from individualist processes within the same explanatory model. In contrast, within criminal justice scholarship Liu’s relationism has been used explicitly to formulate paired, cross-cultural hypotheses and to bridge theoretical gaps between Global North and Global South contexts (Liu, 2024; Yu & Liu, 2025), providing a clear template for how bullying research could move from a general “social-ecological” sensibility to a more sharply specified relationist test program.

In sum, the bullying literature provides strong support for the broad relational premise. However, the evidence that specifically validates Liu’s relationism as a distinct explanatory framework remains limited and largely indirect: bullying studies rarely operationalize the theory’s core relationist values (inner-group attachment, face/honor concern, harmony orientation, holistic thinking), nor do they explicitly test the comparison-paradigm prediction that relationist and individualist processes jointly, and differentially, shape bullying across contexts. As a result, relationism should currently be treated as a promising theory-building and cross-cultural specification program. The most informative next step is targeted tests that measure relationist and individualist constructs side by side and examine whether they predict, mediate, or moderate bullying trajectories in multi-wave, cross-cultural network designs and group-level interventions.

3.2.9 Situational Action Theory (SAT)

Situational Action Theory, originally developed in criminology by Wikström, offers a framework for understanding bullying as a function of personal morality and environmental context intersecting in the moment of action. SAT’s core premise is that whether an adolescent engages in bullying is fundamentally determined by a “moral filter”: their own moral norms about aggression, in combination with the moral norms of the immediate setting, dictate whether bullying is perceived as an acceptable action (Liu et al., 2020; Wikström, 2020; Wikström et al., 2010). According to SAT, a teenager with a strong moral rule that “hurting others is wrong” will be unlikely to bully even if provoked, because the act does not pass their internal moral filter (Caravita et al., 2012; Nickerson et al., 2022). By contrast, a youth with weak moral commitment or who can rationalize aggression is more prone to see bullying as an option. Crucially, SAT argues that traditional “control” factors like self-control abilities or fear of punishment only come into play conditional on the moral context (Hirtenlehner & Leitgöb, 2024; Kammigan, 2023; Wikström et al., 2010). This yields a conditional control proposition: self-control and deterrence effectiveness depend on the presence of a moral conflict. If a student has no moral qualms about bullying, then whether they have high or low self-control might not matter, they may bully simply because it aligns with their goals and they feel little guilt (Antonaccio & Tittle, 2008; Svensson et al., 2010). Conversely, if a student is strongly morally opposed to bullying, external opportunities or weak deterrence are moot, they won’t bully because it’s against their values (Wikström et al., 2010). Only in intermediate cases, for example, a student who knows bullying is wrong but is tempted or pressured in a situation, do self-control and deterrence exert a crucial influence. SAT also predicts that the same propensity will not translate into bullying uniformly across school space and time: exposure to criminogenic settings should amplify the behavioral expression of low morality or weak control, producing propensity × exposure interactions (Farmer et al., 2017; Nickerson et al., 2022; Wikström, 2020; Wikström et al., 2010; Wikström et al., 2018). These propositions are empirically falsifiable: SAT would be challenged if moral norms show little relation to bullying, if predicted interactions fail to appear under adequate measurement of exposure, or if within-person changes in moral judgments and situational opportunities do not covary with bullying over time.

At Level 1, the bullying literature offers several convergent but not definitive patterns that map onto SAT, while also highlighting plausible routes for partial falsification. First, moral cognition is robustly associated with aggressive and bullying-relevant outcomes across school-age samples: meta-analytic work links higher moral disengagement to higher aggression broadly, consistent with SAT’s claim that weakened moral restraint makes harming others more actionable (Bandura et al., 1996; Gini et al., 2014; Killer et al., 2019). In bullying-specific studies, individual moral disengagement and moral emotions show opposite associations with bullying versus defending, suggesting that what adolescents think they can “justify” and what they would feel (guilt/sympathy) cohere with whether bullying is treated as an acceptable option (Thornberg et al., 2015). Importantly, these moral patterns are not only individual: class-level moral disengagement is associated with bullying, aligning with SAT’s emphasis on setting norms as part of the moral filter (Pozzoli et al., 2012). Second, SAT’s conditional-control and propensity × exposure ideas receive partial support from cross-sectional tests showing that low self-control and opportunity co-occur with bullying involvement, implying that the same propensity becomes more consequential when the situation affords bullying (Holt et al., 2014; Moon & Alarid, 2014). Third, bullying-focused SAT applications often find morality to be the most proximal correlate, with control mechanisms (self-control, deterrence) contributing less uniformly, an empirical pattern compatible with SAT’s claim that control primarily regulates behavior under moral conflict, not when the act is already morally neutralized. Yet the same cross-sectional evidence also motivates a more cautious interpretation: moral disengagement measures may partly index post-hoc rationalizations that stabilize self-image after bullying rather than antecedent moral weakness, and punitive climates or perceived sanction risk can be endogenous to prior behavior (Hsieh et al., 2023; Stafford & Warr, 1993; Thornberg et al., 2019). In other words, Level-1 findings are consistent with SAT, but they do not, by design, exclude reverse causality or omitted situational confounding (Pozzoli et al., 2012; Thornberg et al., 2015). A balanced reading is therefore that cross-sectional evidence supports SAT’s core moral architecture while leaving open whether “morality → bullying” is the dominant arrow for all youth, all forms of bullying, and all school ecologies.

Level 2 is still comparatively limited for bullying-specific, full-model tests of SAT, but it sharpens the directionality question that is central to falsification. Some prospective work supports the SAT-consistent claim that moral deficiencies precede bullying trajectories: in a parallel-process latent growth framework, moral deficiencies predicted later bullying involvement, whereas the reverse pathway was not supported in that dataset (Killer et al., 2019; Teng et al., 2020; Wang et al., 2017). At the same time, other longitudinal evidence directly complicates a simple one-way account: a multilevel study of Swedish elementary classes found bidirectional associations such that bullying predicted later moral disengagement and moral disengagement predicted later bullying, suggesting feedback loops in which engaging in bullying may erode moral restraint, which then further facilitates bullying (Thornberg et al., 2019). Longitudinal multilevel work also indicates that classroom norms matter beyond individual differences: classroom-level processes and norms contribute to explaining bullying behaviors over time, consistent with SAT’s insistence that exposure to a moralized setting is not reducible to individual propensity alone (Sentse et al., 2015). Taken together, current Tier-2 evidence supports SAT’s central emphasis on morality and context, but it also implies that bullying may alter the moral filter it is supposed to pass through, introducing dynamic endogeneity that static cross-sectional designs cannot adjudicate (Thornberg et al., 2019; Wikström, 2020). Progress toward a more decisive test likely requires designs that measure situational exposure with higher temporal resolution and separate within-person change from between-person differences, an agenda echoed in methodological syntheses of SAT testing and illustrated by experimental approaches that manipulate situational inputs to observe rule-breaking choices (Hardie & Rose, 2025; Sentse et al., 2015; Wikström, 2020; Wikström et al., 2018).

SAT is best characterized at present as a strongly supported morality-centered framework for bullying, while its full situational-interaction model awaits more decisive multi-wave tests that jointly measure moral propensity, setting-level moral climates, and situational exposure at sufficient temporal resolution to adjudicate directionality and interaction claims.

3.2.10 Routine Activity Theory (RAT)

Routine Activity Theory, a criminological framework, explains harmful events as the situational convergence of a motivated offender, a suitable target, and the absence of capable guardianship (Cohen & Felson, 1979; Hollis-Peel et al., 2011). Applied to bullying, RAT shifts attention from stable “bully” and “victim” traits to the opportunity structures embedded in everyday school routines: predictable transition times, semi-public spaces, and moments when adult attention is thin. This emphasis is compatible with evidence that bullying is a pervasive developmental stressor with enduring mental-health and educational sequelae, which makes its situational contingencies a consequential target for theory and prevention (Arseneault, 2018; Copeland et al., 2013; Modecki et al., 2014). In criminology, routine activities are not merely background conditions; they are mechanisms that shape exposure to potential offenders and the availability of guardianship in everyday life (Osgood et al., 1996). In schools, this logic foregrounds “unowned” spaces and times, corridors, cafeterias, buses, and other transition zones, where the social organization of supervision may be ambiguous, even when the institution is formally rule-bound (Astor et al., 1999). RAT therefore yields falsifiable predictions: bullying should concentrate in recurrent low-guardianship micro-settings; manipulating guardianship or structure should shift rates; and if opportunity structures change while bullying does not (or vice versa), the explanatory scope of RAT is bounded.

Most evidence bearing on RAT in bullying research remains Tier 1, consisting of cross-sectional and descriptive studies that map bullying to routines, places, and supervision. Naturalistic observations show that bullying episodes occur more often in playground contexts than in classrooms and that adult intervention is limited, consistent with the idea that guardianship varies systematically across school settings (Craig et al., 2000; Craig & Pepler, 1998; Lynn Hawkins et al., 2001). Students’ spatial accounts of school aggression highlight transition areas and times with limited teacher presence, consistent with uneven guardianship across the school day (Astor et al., 1999). Survey analyses that explicitly invoke routine-activity mechanisms link bullying victimization to variation in school security and routine involvements (Cecen-Celik & Keith, 2016; Hong et al., 2024). Yet, these correlational patterns are not uniquely diagnostic. In some studies, physical security measures and other “hard” safety strategies show weak, null, or heterogeneous associations with peer victimization, which could mean that such measures are poor proxies for “capable guardians,” or that they are deployed reactively in high-risk schools (Blosnich & Bossarte, 2011; Gerlinger & Wo, 2016). Perceptual gaps between students and staff further complicate inference about guardianship, because staff may underestimate bullying and misjudge its locations. Finally, selection and feedback are plausible: perpetrators may seek unsupervised contexts, and frequent bullying may itself erode perceived guardianship or reshape routine participation, blurring cause and consequence (Blosnich & Bossarte, 2011; Cecen-Celik & Keith, 2016).

Tier 2 evidence is less common for routine-activity constructs in bullying, but it underscores why directionality matters. In longitudinal analyses of U.S. adolescents, higher or increasing levels of school security measures have been linked to higher reports of some victimization indicators over time, a pattern consistent with reactive deployment and confounding by baseline risk rather than straightforward deterrence (Fisher & Widdowson, 2023). This kind of finding does not falsify RAT so much as it exposes a measurement and identification problem: “guardianship” is endogenous and may not be well captured by counts of cameras, guards, or metal detectors. The strongest evidence that opportunity structures matter often comes indirectly from whole-school intervention evaluations and meta-analyses. Classic program evaluations show that comprehensive school-based interventions can reduce bullying, and such programs commonly include situational components, clear rules, increased supervision, and structured routines, that map plausibly onto RAT mechanisms (Olweus, 1993). Meta-analytic syntheses likewise conclude that school-based programs reduce bullying perpetration and victimization on average, although component-level causal attribution remains difficult (Gaffney et al., 2021; Ttofi & Farrington, 2011). A further Tier 2 implication concerns potential displacement: opportunity reduction could shift aggression to alternative settings rather than eliminate it, yet systematic reviews of situational prevention suggest displacement is far from inevitable and that diffusion of preventive benefits is at least as plausible.

Rare Tier 5 evidence also exists: in a 2-year cluster-randomized trial across 16 primary schools, Farmer et al. (2017) experimentally altered recess opportunity structures by redesigning playground environments (loose parts, reduced rules, and increased risk/challenge) and tracked bullying indicators over time. Although effects were mixed, most notably higher reports of pushing/shoving, the design demonstrates that RAT-style opportunity mechanisms can be tested experimentally, instead of inferred from cross-sectional place mappings.

Overall, RAT is best read as a high-resolution account of how bullying is patterned by school routines and micro-settings, with a causal status that remains contingent. More decisive tests require designs that separate prevention from response and measure opportunity structures at the level RAT specifies. Natural experiments induced by scheduling reforms or supervision reallocations, combined with event-based spatiotemporal records, could test whether guardianship shifts precede bullying shifts rather than follow them. RAT also implies that opportunity structures can span offline and online contexts: cyber and traditional bullying are intertwined, and online settings may function as low-guardianship arenas. This motivates explicit tests of displacement versus diffusion across settings, rather than assuming either outcome. Finally, because guardianship in schools likely has relational components in addition to surveillance, research should distinguish interactionist guardianship from security hardware and examine their distinct links to victimization.

3.3 Cross-Theoretical Synthesis: Emerging Conclusions

The foregoing review of theories and their evidentiary status reveals that the knowledge base on bullying’s causes is uneven across theoretical domains. This unevenness should not be mistaken as simple confirmation that some theories are “correct” and others “wrong”. Rather, it reflects that certain theories have been easier to probe with available data, while others posit causal structures that typical study designs have not adequately captured (Hamaker et al., 2015; Maxwell & Cole, 2007). A recurring theme is that much of the empirical literature, historically, has relied on cross-sectional associations, linking individual traits or perceptions with bullying involvement at a single time. These correlations often satisfy predictions of multiple theories simultaneously since many risk factors co-occur, but they rarely satisfy strict causal criteria (Arseneault, 2018). A limitation seen across many studies is that temporality is assumed rather than demonstrated. Researchers often interpret traits as causes and bullying as the effect, but without longitudinal data or experiments, the direction of influence can be ambiguous, e.g., does low empathy lead to bullying, or could bullying behavior erode empathy. This issue indicates a saturation at Tier 1 in the field, a plethora of risk factor correlations, with slower progress into Tier 2+ where temporal ordering and conditional relationships are tested (Arseneault, 2018; Smith, 2016; Volk et al., 2014).

Another cross-cutting complication is selection bias and omitted variables, especially involving peer networks and contextual factors. The social nature of bullying introduces self-selection dynamics: for example, aggressive youths may cluster together, which boosts the apparent effect of “peer influence” on each other, when part of it was that they chose those peers (Manski, 1993; McPherson et al., 2001). Similarly, victims might change classrooms or schools, confounding attempts to assess school-level influences. These sorts of issues undermine straightforward causal inference in theories like social learning, collective efficacy, or relationism at Tier 2, because any longitudinal change could as easily be due to who enters or exits the peer group as due to genuine influence (Shalizi & Thomas, 2011; Steglich et al., 2010). Time-varying confounds, shifting family situations, developmental changes, evolving school climates, are likely the norm in adolescent bullying (Espelage et al., 2014). For example, a student might start bullying because of a parental divorce that coincides with entering middle school, disentangling the cause, family ACE vs. school environment vs. developmental stage, is enormously challenging without specialized designs. Standard regression controls often fail to capture these dynamic, multi-level confounds adequately (Cole & Hernán, 2008; Greenland, 2017; Robins et al., 2000).

Perhaps the most fundamental challenge illuminated by our review is interference and non-independence: by definition, bullying involves at least two people, and often affects many bystanders. This means a simple individual-level causal story is inherently incomplete (Lynn Hawkins et al., 2001; Salmivalli, 2010; Swearer & Hymel, 2015). If one student stops bullying, that could change the likelihood that another student becomes a bully or a victim, because peer group dynamics shift. Traditional analytic methods treat such spillover as a problem, but in bullying research this interdependence is actually central to the causal process. Many “individual” risk factors may operate by shaping how a youth interacts with others and how others respond, thus, a purely individual-centric analysis can mislead (Faris & Felmlee, 2011). In our theory-by-theory survey, we saw repeatedly that the counterfactual for one child, that is what would happen if they didn’t have a certain risk factor, often depends on others’ behavior. This kind of interaction, person × peer, is exactly what creates ambiguity in testing theories: many studies did not explicitly model peers or context, so their findings can’t easily distinguish whether, say, “low self-control” mattered on its own or only in a permissive environment (Mathur & VanderWeele, 2020; Moon & Alarid, 2014; Moon et al., 2008). In short, interference and unmodeled context can produce false confidence, one might attribute bullying to an individual trait when in reality that trait only mattered because of an unobserved social situation.

Given these complexities, we can assert with confidence that bullying is a multi-causal behavior with contributions from individual dispositions. All the theories reviewed capture part of this picture, and none is entirely negated by existing evidence. The field has made progress in identifying risk factors and correlates. However, what we can causally claim is narrower. Thanks to some longitudinal and experimental studies, we have credible evidence for a few causal relationships: for instance, well-implemented whole-school anti-bullying programs can causally reduce bullying by 20-30% (Ttofi & Farrington, 2011). This implies that aspects of several theories including increased guardianship (RAT), changes in group norms (relationism/social learning), when done together, do cause declines in bullying. Additionally, studies with stronger designs suggest that peer context is not just noise but a causal force: e.g., the Paluck et al. (2016) experiment showed that leveraging peer influence by changing a few students’ behavior led to broader reductions in conflict, demonstrating causally that peer dynamics drive bullying rates. Thus, we can credibly claim that interventions targeting social systems (peers, classrooms, schools) are necessary for meaningful bullying reduction, an assertion supported by both theory and experimental evidence. Overall, the literature indicates that to explain bullying, no single theory is sufficient, bullying is a result of risk accumulating across multiple domains, and preventive efforts falter when they ignore that complexity (Hong & Espelage, 2012; Swearer & Hymel, 2015).

This state of affairs points to a clear methodological agenda for the field: the priority should be shifting research from cataloguing more Tier-1 correlations to executing Tier 3-5 designs that isolate mechanisms. Future studies would do well to emulate the integrative approaches we discussed, for example, using natural experiments like policy changes, employing network-based randomized interventions, and analyzing within-person changes over time (Gerlinger et al., 2021; Paluck et al., 2016). By explicitly designing studies that approximate counterfactuals, we can start to adjudicate between theories that currently all receive some support in correlational form. Given that bullying intrinsically involves interactions, interdisciplinary and multi-method approaches may be required to capture the full causal web (Robins et al., 2000; Salmivalli et al., 1996).

In conclusion, we can credibly claim that we know a good deal about who is at risk of bullying involvement and under what conditions bullying tends to flourish; yet, our understanding of causal pathways is still evolving.

4. The empirical map: prevalence patterns and form distributions

4.1 A global snapshot—and why “global prevalence” is an oversimplification

The empirical evidence reveals that bullying is a widespread phenomenon among youth worldwide, but its prevalence varies markedly across regions and studies. Large international surveys indicate that roughly 20-30% of adolescents experience some form of bullying involvement, although estimates differ depending on measurement criteria (Ariani et al., 2025; Hosozawa et al., 2021; Menesini & Salmivalli, 2017). For example, one meta-analysis of global studies found about 25% of children and adolescents have been victims of bullying, with around 16% engaging as perpetrators, a substantial minority across diverse societies (Ariani et al., 2025). Another cross-national survey of over 300,000 students in 83 countries reported a pooled 30.5% prevalence of being bullied in the past month, but with extremely wide regional differences (Biswas et al., 2020; Craig et al., 2009).

The highest rates were observed in parts of the Global South: in the World Health Organization’s Eastern Mediterranean and African regions, approximately 40-45% of adolescents reported recent bullying victimization, compared to under 10% in some European countries (Biswas et al., 2020). Such disparities underscore that bullying is a near-universal issue, yet one that manifests to very different degrees across socio-cultural contexts. They also hint that reported prevalence is contingent on many factors, including not only real behavioral differences, but also methodological aspects like how bullying is defined and measured in each study (Hosozawa et al., 2021; Menesini & Salmivalli, 2017; Smith, 2016). In short, bullying affects youth in every world region, but any “global” rate is an oversimplification: beneath the overall figure lie pronounced regional patterns and inconsistencies arising from contextual influences and survey approaches.

4.2 Western/high-income settings: moderate prevalence, with stability or declines in some countries

Regional prevalence patterns observed in research suggest a rough divide between many high-income Western countries and certain low- and middle-income regions, albeit with important exceptions. In much of Europe, North America, and other Western settings, studies tend to find moderate rates of bullying by global standards. National surveys in these countries often report that roughly 10-20% of adolescents are regularly victimized(Menesini & Salmivalli, 2017; Nansel et al., 2001; Smith, 2016). For instance, the Health Behavior in School-aged Children (HBSC) survey in European and North American countries has consistently found around 11% of students bullied at school on a frequent basis in recent years (Craig et al., 2009; Menesini & Salmivalli, 2017). In some Northern European nations, the prevalence is even lower, historically among the lowest in cross-national comparisons (Biswas et al., 2020; Chester et al., 2015).

Trends in many Western countries have been stable or declining over time, likely reflecting sustained prevention efforts and heightened awareness. A comparative analysis of 33 mainly Western countries showed significant decreases in self-reported bullying victimization from 2002 to 2010 in about one-third of those countries, and no widespread increases in most others. These encouraging trends suggest that, at least in settings with long-standing anti-bullying initiatives and robust research, school bullying may be gradually becoming less common (Chester et al., 2015; Molcho et al., 2009). Nevertheless, even in the best-case examples, a non-trivial minority of youth still report being bullied, indicating no society has eliminated the problem (Craig et al., 2009; Hosozawa et al., 2021; Nansel et al., 2001). Moreover, Western countries themselves are not monolithic: there are intra-regional differences, e.g. high er rates in some Eastern European and Mediterranean countries than in Northern Europe, that remind us culture and policy matter (Biswas et al., 2020; Craig et al., 2009; Due et al., 2005; Sittichai & Smith, 2015). Overall, however, the Western experience of bullying is characterized by moderate prevalence that has plateaued or declined in many cases, potentially due to decades of research-based interventions and a normative shift that increasingly condemns school violence.

4.3 Global South: higher reported prevalence, but with sparse evidence and greater measurement uncertainty

Surveys conducted in parts of the Global South often reveal higher reported bullying rates, though data from these regions are comparatively sparse. Several low- and middle-income countries in Asia, Africa, the Middle East, and Latin America show alarmingly high levels of youth bullying in the available studies (Biswas et al., 2020; Sittichai & Smith, 2015). As noted, the Eastern Mediterranean and African regions have the highest prevalence according to the GSHS-based analysis, over 40% monthly bullying victimization (Biswas et al., 2020). Individual country reports echo this: for instance, survey data have indicated that in some Sub-Saharan African countries, well over half of adolescents have experienced bullying recently (Fleming & Jacobsen, 2010; Smith et al., 2023). Similarly, studies in certain Central and South American countries and parts of the Caribbean often document bullying rates on the order of 30-50%, significantly above the global average (Delprato et al., 2017; Hosozawa et al., 2021; Modecki et al., 2014).

These high figures might reflect genuine differences in school environments, for example, larger class sizes, fewer resources, or greater exposure to community violence and socio-economic stressors, which can all exacerbate peer aggression (Davis et al., 2020; Elgar et al., 2009). At the same time, it is crucial to acknowledge that measurement differences and cultural norms might inflate the reported rates in some developing contexts (Modecki et al., 2014; Solberg & Olweus, 2003). In societies where adult supervision in schools is limited or corporal punishment is relatively common, youth might interpret survey questions about “being bullied” more broadly to include one-off aggressive incidents or teacher aggression, leading to higher affirmative responses (Sittichai & Smith, 2015; Smith, 2016).

Nonetheless, even with cautious interpretation, the existing data signal that bullying is often highly prevalent in the Global South. This pattern is frequently accompanied by risk factors such as low levels of peer or parental support, which are more prevalent in those contexts and are strongly associated with victimization. For example, one study found adolescents in low- and lower-middle income countries were more likely to be bullied if they reported weak family understanding or friend support, suggesting that social support deficits may partly mediate the high bullying rates observed in those settings (Biswas et al., 2020; Bowes et al., 2009; Due et al., 2005). In sum, many developing regions face a significant burden of school bullying, though the evidence base in these areas remains less extensive. The contrast between generally lower rates in Western countries and higher rates in parts of the Global South underscores the influence of structural and cultural factors, as well as the need for more research in under-studied locales (Elgar et al., 2009; Smith, 2016).

4.4 Asia is not a single case: heterogeneity within and across subregions

Within Asia, bullying patterns are heterogeneous and illustrate how regional averages can mask internal diversity. East Asian countries, for instance, have drawn attention both for relatively low official rates of bullying in surveys and for infamous cases of severe bullying and bullying-related suicides that suggest an iceberg effect (Akiba, 2004; Koyanagi et al., 2019). According to a study of 71 countries, South Korea had one of the lowest rates of frequent bullying victimization, only about 9% of 15-year-olds there reported being bullied “more than a few times a month,” compared to a global mean of 30% (Craig et al., 2009; Elgar et al., 2009; Hosozawa et al., 2021). Similarly, national surveys in Japan have often reported bullying prevalence at the lower end of international comparisons (Osuka et al., 2019). These East Asian contexts tend to have highly structured school systems and cultural norms emphasizing group harmony, which might suppress open physical bullying but can give rise to more indirect or covert forms. Indeed, qualitative accounts from Japan and Korea have described bullying that centers on social exclusion and rumor-spreading, aligning with a cultural tendency to avoid direct confrontation (Koo et al., 2008; Olweus, 2013).

In contrast, some countries in Southeast Asia report much higher bullying involvement. For example, the Philippines has been highlighted as having among the highest bullying rates globally: approximately 65% of Filipino adolescents in one international survey said they were bullied, the highest proportion recorded (Hosozawa et al., 2021). Other ASEAN nations like Malaysia, Indonesia, and Thailand typically report prevalence levels well above the Western average (Pengpid & Peltzer, 2019; Tan et al., 2019). A review by Sittichai and Smith (2015) noted that in many Southeast Asian studies, verbal bullying was the most commonly reported form, and overall victimization rates ranged widely but were often elevated compared to Western benchmarks.

These differences within Asia underscore that cultural and socio-political contexts, such as the disciplinary style of schools, the value placed on hierarchy, and even the way “bullying” is linguistically framed, can lead to divergent bullying experiences. It is telling that even within East Asia, where cultures share certain features, countries like Korea and Japan show relatively low survey-reported rates whereas places like China and Hong Kong have reported moderate rates, and some Southeast Asian locales are higher (Hosozawa et al., 2021; Rappleye & Komatsu, 2020; Sittichai & Smith, 2015; Xing et al., 2022). Such variability cautions against treating “Asia” as a single case and reinforces that local conditions, from school discipline policies to social media usage patterns, critically shape bullying dynamics. In Asian regions with strong academic pressure and strict school discipline, e.g., East Asia, bullying may take more subtle or group-based forms that sometimes elude standard measures. Meanwhile, in regions undergoing rapid social change or where youth supervision is inconsistent, bullying may be more frequent or visible (Biswas et al., 2020; Nguyen et al., 2019; Wang et al., 2020). Ultimately, Asia’s mixed profile contributes to the broader empirical map by demonstrating how economic development, cultural norms, and educational systems intersect to produce different bullying outcomes, even among geographically proximate societies (Hosozawa et al., 2021; Sittichai & Smith, 2015; Smith, 2016).

4.5 Beyond geography: form distributions, gender patterns, and the online–offline continuum

4.5.1 Cross-cultural regularities and variations in bullying forms

Beyond prevalence rates, there are noteworthy cross-cultural consistencies and variations in the forms of bullying and the profiles of those involved. Research across many countries finds that verbal bullying is the most common form of bullying reported by students (Hosozawa et al., 2021; Li et al., 2020; Menesini & Salmivalli, 2017; Sittichai & Smith, 2015). For instance, in surveys from Europe to Asia, a plurality of victims often report being taunted or made fun of more than any other type of aggression (Sittichai & Smith, 2015; Wang et al., 2009). Social or relational bullying is also prevalent across cultures, frequently ranking a close second to verbal harassment in incidence (Hosozawa et al., 2021; Smith, 2016). In contrast, physical bullying tends to be less common than verbal or relational forms in most student populations, especially by mid-adolescence (Craig et al., 2009; Hosozawa et al., 2021; Menesini & Salmivalli, 2017). A recent analysis of 71 countries found that on average about 21% of 15-year-olds had experienced verbal bullying and 20% relational bullying, compared to 15% for physical bullying, indicating that non-physical tactics slightly predominate worldwide.

This pattern likely reflects a developmental shift: physical aggression is more frequent in childhood and tends to decline with age, whereas older adolescents increasingly engage in more sophisticated social forms of aggression (Card et al., 2008; Pepler et al., 2008). One cross-national study spanning 40 countries observed that younger adolescents faced more direct physical and verbal attacks, whereas older teens experienced more indirect bullying, illustrating this age-linked shift in form (Craig et al., 2009; Marsh et al., 2011). Culturally, there may be differences in emphasis, for example, some East Asian contexts place particular weight on social exclusion as a harmful tactic. However, empirical comparisons suggest that actual frequencies of exclusion vs. other forms do not vary as drastically as the cultural narratives imply (Smith, 2016). The modal pattern is that bullies choose tactics that are effective and acceptable within their context, whether that means punches on the playground or shunning on social media, but the underlying intent to harm and exert power is a common thread (Menesini & Salmivalli, 2017; Volk et al., 2014).

4.5.2 Gender patterns: broad regularities with context-specific nuances

Gender differences in bullying follow some universal trends with contextual nuances (Biswas et al., 2020; Craig et al., 2009; Menesini & Salmivalli, 2017). The GSHS data showed males had greater odds of being victimized than females in most countries (overall OR -1.2 for boys) (Biswas et al., 2020). Boys also engage more in physical fights and direct aggression according to international surveys. Girls, on the other hand, are often thought to experience more relational bullying, and indeed some studies find smaller gender gaps or even slightly higher victimization for girls when indirect forms are considered (Card et al., 2008; de Looze et al., 2019). In 29 of 40 countries in one HBSC analysis, girls reported higher victimization rates than boys, even though boys reported bullying others more, suggesting that girls may be frequently targeted in ways not captured by narrow measures of overt aggression (Craig et al., 2009).

However, the gender differences are not uniform across all contexts; cultural norms about gender behavior can influence both the expression of bullying and the willingness to report it (Hosozawa et al., 2021; Menesini & Salmivalli, 2017). In more patriarchal or traditional societies, girls might under-report bullying, especially sexual or physical harassment, or their victimization may be less visible, whereas in societies with greater gender equality, the reported gap between boys and girls often narrows (Carbone-Lopez et al., 2010; Cosma et al., 2022; Hosozawa et al., 2021). Overall, though, the power imbalance inherent in bullying can manifest through gendered patterns: boys more often leverage physical power, while girls might leverage social networks to ostracize or demean (Menesini & Salmivalli, 2017). Despite these trends, it is crucial to avoid stereotypes, both genders participate in all forms of bullying to some extent, and both suffer harm from victimization. Any effective anti-bullying strategy must therefore be attuned to the forms of bullying most salient to each gender in the given cultural context.

4.5.3 Cyberbullying: a growing, cross-border layer shaped by digital infrastructure and governance

In the past two decades, the rapid expansion of youth access to the internet and smartphones has given rise to cyberbullying as a new and borderless form of aggression, and researchers have been mapping its prevalence alongside traditional bullying (Aboujaoude et al., 2015; Livingstone & Smith, 2014). Broadly, cyberbullying, online harassment, denigration, or social exclusion via digital technologies, remains less common than traditional face-to-face bullying, but the gap is closing (Li et al., 2024; Modecki et al., 2014). However, cyberbullying rates are rising in many regions as internet connectivity grows. Upward trends in online bullying have been documented in middle-income countries as more youth come online via social media (Aboujaoude et al., 2015). It appears that digital platforms are amplifying peer victimization, often extending the reach of bullying beyond school hours and geographic boundaries (Kowalski et al., 2014; Kwan & Skoric, 2013; Menesini & Salmivalli, 2017). Notably, there is a large overlap between offline and online bullying: the majority of cyberbullying victims are also bullied in person, and vice versa, rather than representing entirely distinct groups. This overlap suggests that cyberbullying is often another arena in which the same power imbalances play out (Cosma et al., 2022; Kwan & Skoric, 2013; Li et al., 2024; Modecki et al., 2014).

Nevertheless, cyberbullying poses unique challenges as it can occur anonymously and virally, potentially leading to severe psychological impacts (Gini et al., 2018; Li et al., 2024). The emergence of cyberbullying has added a new dimension to the global profile of bullying: it is a growing problem everywhere digital technology is pervasive, although current evidence still indicates it has not surpassed traditional bullying in frequency. The form distribution of bullying now must encompass this online-offline continuum, which varies with a country’s digital infrastructure and youth online behavior (Li et al., 2024; Livingstone & Smith, 2014; Menesini & Salmivalli, 2017). Thus, any empirical map of bullying in the 2020s is incomplete without tracking cyber-mediated abuse alongside schoolyard encounters.

4.6 Linking the empirical map to theory: what may generalize, what is likely context-dependent

These descriptive patterns naturally raise a cautious interpretive question: which theoretical propositions appear to travel well across contexts, and which likely hinge on institutional, cultural, and governance conditions? The goal here is to use the empirical map to clarify where “universality” is plausible and where context sensitivity is expected. A further caution is epistemic: what is treated as “universal” in bullying theory often reflects where theory is produced and which contexts set the default questions. Asian criminology and southern criminology explicitly problematize this Global North default and thus motivate treating theoretical portability as an empirical question, one that should be tested and, where needed, re-specified through two-way dialogue across North, Asia, and the Global South (Carrington et al., 2016; Liu, 2017, 2021).

More plausibly context-dependent propositions (sensitive to institutions and environments) are as follows. Cross-national differences in bullying prevalence and dominant forms are difficult to explain without explicitly embedding adolescents in the institutional and environmental “containers” that shape everyday interaction—families, schools, communities, and macro-level social conditions (Clayton R Cook et al., 2010; Juvonen & Graham, 2014; Swearer et al., 2010). From this perspective, variation across countries plausibly reflects differences in inequality, exposure to community violence, and school-system norms that structure supervision, competition, and peer regulation, thereby shaping both how often bullying occurs and which forms are most viable or salient (Elgar et al., 2009; Smith, 2016). Policy regimes and disciplinary systems add another layer of contingency: whether punitive approaches deter bullying or backfire (including displacement into covert or cyber forms) likely depends on how discipline is implemented, perceived legitimacy, and the extent to which labeling dynamics are triggered locally (Hall, 2017; Strindberg et al., 2020). In the digital domain, cyberbullying is especially sensitive to platform affordances and the surrounding governance ecology, moderation practices, reporting infrastructure, and state or institutional responses, so ostensibly similar online behaviors may translate into different prevalence and harm profiles under different regulatory environments (Kowalski et al., 2014; Sittichai & Smith, 2015).

When it comes to potentially more “universal” propositions (broad cross-context alignment), across diverse contexts, several regularities recur strongly enough to plausibly function as cross-cultural anchors for theory: bullying typically involves an abuse of power and often operates as a strategy for dominance or status within peer hierarchies, consistent with social dominance and evolutionary accounts of adolescent status competition (Hawley, 1999; Volk et al., 2014). A second robust alignment is that bullying involvement is consistently linked to psychosocial harm, including worse health, emotional well-being, and academic functioning, as repeatedly observed in cross-national evidence (Arseneault, 2018; Gini et al., 2018; Juvonen & Graham, 2014; Nansel et al., 2001). Third, bullying is rarely sustained by dyads alone: peer-group processes and social reinforcement matter across settings, with bystander attention, approval, and group dynamics functioning as a stabilizing mechanism in both Western and non-Western research traditions (O'Connell et al., 1999; Rambaran et al., 2020; Salmivalli, 2010). Fourth, pathways consistent with learned aggression appear broadly relevant: exposure to violence in home, community, or media environments is repeatedly associated with bullying perpetration, cohering with social learning accounts even as causal identification remains methodologically challenging (Hong & Espelage, 2012). Finally, there is evidence for a recurring developmental regularity in form distribution: as children age, bullying tends to shift from more direct physical tactics toward more indirect relational strategies, a pattern documented across multiple national contexts (Björkqvist et al., 1992; Menesini & Salmivalli, 2017; Wolke et al., 2013). Taken together, these propositions suggest that while bullying’s intensity and surface expression may vary widely by context, its core logic, power asymmetry, peer reinforcement, and predictable developmental shifts, alongside reliably harmful consequences, shows substantial alignment across societies.

In short, the empirical map is consistent with a two-layer view: bullying’s core relational features and many micro-level processes show broad cross-cultural recurrence, whereas between-country differences in prevalence and the relative salience of forms likely depend on context-specific institutional, cultural, and governance factors.

4.7 Cognitive inequality in the evidence base—and a forward agenda for a genuinely global map

4.7.1 Cognitive inequality: when “global theories” become global by default

A critical insight emerging from this review is the existence of a knowledge imbalance or cognitive inequality in bullying research (Slonje et al., 2025; Zych et al., 2015). Theoretical frameworks have often been presumed global in applicability, yet many are derived from evidence in a relatively narrow band of contexts, predominantly Western, educated, industrialized, rich, and democratic, the so-called WEIRD societies (Henrich et al., 2010). Asian criminology argues that non-Western contexts, especially Asia, should not be treated as one-way recipients of Northern concepts; instead, they can serve as innovative sites for revising, extending, and generating criminological theory through an Asian criminological paradigm that links global North and South in two-way dialogue (Liu, 2009, 2017, 2021; Zhang & Liu, 2023). Southern criminology makes a parallel claim for the Global South, highlighting the power relations embedded in the hierarchical production of criminological knowledge and calling for decolonizing and democratizing the theoretical toolbox rather than exporting it as a universal template (Carrington et al., 2018; Carrington et al., 2016; Dimou, 2021).

Seen through this lens, the bullying evidence base exhibits clear cognitive inequality. Regions such as sub-Saharan Africa, parts of South Asia, the Middle East, and Southeast Asia have comparatively few longitudinal or experimental studies on bullying (Sittichai & Smith, 2015; Smith, 2016). As a result, concepts like “effective anti-bullying strategies” or assumptions about bully/victim role profiles are sometimes generalized worldwide without direct verification in diverse settings.

This amounts to a form of epistemic inequity: the experiences of youth in well-studied contexts (e.g., North America, Europe, East Asia) carry disproportionate weight in shaping “universal” theories, while youth in less-studied contexts are underrepresented in the data that inform theory (Henrich et al., 2010; Sittichai & Smith, 2015). In the bullying domain, the uncritical universalization of Global North frameworks can even resemble a subtle form of epistemic bullying: it normalizes Northern-derived concepts as default, casts Southern and Asian realities as deviations to be “explained away,” and can marginalize locally grounded scholarship in the very regions where the evidence base is thinnest (Carrington et al., 2016; Liu, 2021).

In essence, the current state of research often elevates theories developed in well-researched settings to a “global” status, only because alternate perspectives from elsewhere have not been available to challenge or refine them instead of being proven globally. The recognition calls for a more inclusive research agenda to correct the imbalance.

4.7.2 A concrete agenda: comparable measurement, sentinel cohorts, policy-evaluation partnerships, and heterogeneity-aware causal learning

Looking ahead, advancing the empirical map of bullying and mitigating these knowledge gaps will require deliberate strategies. We propose an agenda with several key components to develop a more globally informed understanding of bullying.

First, we need a small, shared measurement “core” that researchers can drop into surveys without rebuilding instruments from scratch (Menesini & Salmivalli, 2017; Smith, 2016). Right now, prevalence shifts depending on whether questionnaires use the word “bullying,” the recall window, and what counts as repetition (Solberg & Olweus, 2003). A practical fix is a short module that (a) starts with the same plain-language definition, like intent, power imbalance, repetition or sustained threat, (b) asks about specific behaviors in a fixed recall window, and (c) uses the same response scale and cut-points. Translation should follow back-translation plus cognitive interviewing, and teams should report basic invariance checks rather than assuming equivalence. The core can be complemented by optional local items, but the shared block should remain unchanged. CDC and UNESCO definitions are a start; the next step is routine adoption and a public codebook so “bullying” means the same thing across surveys (Cascardi et al., 2014; Cornell & Limber, 2015).

Second, we need sentinel cohorts in places where the map is still thin. Pick a small set of schools or communities in Africa, South Asia, the Middle East, and Latin America, and run the same core bullying module on a fixed schedule (Liu, 2021; Liu & Li, 2024; Liu et al., 2025). Keep the design consistent, same grade bands, same season, same informants, so changes over time are interpretable. Add a short “context panel” (school resources, smartphone/internet access, major policy or platform changes) to explain why trends move. The Global School-based Health Survey could be a practical vehicle if repeated regularly with a stronger bullying block (Biswas et al., 2020). Crucially, sentinel sites should be co-led locally, with budget lines for training, translation, and data stewardship, so the work reduces cognitive inequality (Sittichai & Smith, 2015; Smith, 2016).

Third, treat policy rollouts as built-in experiments (Craig et al., 2008; Craig et al., 2009). Ministries, school districts, and NGOs change practice all the time, like curricula, teacher training, reporting systems, digital safety rules, yet evaluation is often an afterthought. Partnerships can fix that by embedding a simple evaluation plan from day one: a clear roll-out schedule, comparable outcome measurement (the shared bullying core), and basic implementation tracking (who received what, when, and with what dose). When randomization is feasible, use cluster RCTs or stepped-wedge designs; when it is not, use transparent quasi-experimental approaches (e.g., DiD around phased implementation) (Campbell et al., 2012; Hemming et al., 2015). The goal is not just “does it work?” but “what has to change for it to work here?”—and those adaptations should be documented. These partnerships also let researchers evaluate system-level moves, such as new anti-bullying laws or platform moderation changes, using administrative data like attendance, discipline, school transfers to complement self-reports.

Last but not least, we should treat heterogeneity as a first-order result (Athey & Imbens, 2017; Rothwell, 2005). Bullying determinants and interventions rarely work the same way for every student or school, yet many papers report only one average effect. That average can hide high payoff in some contexts and near-zero (or adverse) effects in others (Athey & Imbens, 2017; Ttofi & Farrington, 2011). A workable next step is to pair causal machine-learning tools, especially causal forests, with population attributable fractions (PAFs) to turn heterogeneity into prevention priorities (Benichou, 2001; Rockhill et al., 1998; Susan et al., 2019).

The workflow is practical. Begin with a design that already has causal leverage (a cluster RCT, stepped-wedge roll-out, or a credible policy shock) (Campbell et al., 2012; Hemming et al., 2015). Then use a causal forest to estimate conditional average treatment effects across rich covariates, including baseline bullying, school climate, supervision, digital access, marginalization (Athey & Imbens, 2017; Susan et al., 2019). The point is not to replace design with algorithms, but to let the algorithm search for effect modifiers once the comparison is credible. Next, translate those heterogeneous effects into impact using PAF-style quantities: “What share of bullying cases could be prevented under a defined intervention?” PAFs are routine in epidemiology because they express results as avoidable burden under a specific scenario (Benichou, 2001; Rockhill et al., 1998). In bullying research, the same logic can be applied to modifiable determinants or to interventions (Astor et al., 1999; Kärnä et al., 2011). Future study could compute a universal PAF and, more importantly, a targeted PAF: allocate resources to the schools or students where the causal forest predicts the largest benefit, and report the expected reduction under realistic budget constraints.

This agenda travels well beyond epidemiology. In criminology, it can help target violence-prevention resources; in education, it can identify where climate reforms or digital-safety policies deliver the biggest gains (Athey & Imbens, 2017; Braga et al., 2019; Polanin et al., 2022; Waasdorp et al., 2012). The outputs can be concrete: ranked settings, expected reductions with uncertainty, and transparent assumptions. The guardrails should be concrete too: define interventions clearly, check overlap/positivity, report sensitivity to unmeasured confounding, and be explicit about spillovers (because one student’s treatment can affect others). Done well, causal forests plus PAF move the field from “these factors matter” to “here is what to do, for whom, and how much harm it could plausibly prevent”(Athey & Imbens, 2017; Rockhill et al., 1998; Susan et al., 2019; Wager & Athey, 2018).

5. Discussion on theory evidence and empirical map

5.1 Global significance and purpose: from a global burden to a causal agenda

Bullying in adolescence is a multifaceted global phenomenon, and our review provides a comprehensive synthesis of its determinants through both theoretical and empirical lenses. The gravity of bullying as a worldwide public health issue cannot be overstated: nearly one in three youth experience some form of peer victimization, with prevalence varying dramatically across countries (Biswas et al., 2020; Modecki et al., 2014). For instance, surveys indicate that bullying victimization rates range from under 10% in certain European samples to over 40% in parts of Africa and the Eastern Mediterranean (Biswas et al., 2020). These patterns underscore that while bullying is ubiquitous, its expression and frequency are shaped by cultural, socioeconomic, and policy contexts (Hong & Espelage, 2012; Menesini & Salmivalli, 2017).

Against this backdrop, identifying why bullying arises and persists is of both practical importance and theoretical interest. By mapping out credible causes of bullying, spanning individual propensities, family dynamics, peer group processes, school climates, and broader structural forces, this review moves the field toward a more causally informed understanding of youth aggression (Arseneault, 2018; Clayton R Cook et al., 2010). Our approach, which integrates classic behavioral and criminological theories with a hierarchy of evidence strength, represents an effort to advance bullying research from a descriptive catalogue of correlates to an explanatory framework grounded in causal inference (Dishion & Tipsord, 2011; Ttofi & Farrington, 2011; Walters & Espelage, 2018).

5.2 Evidence-tiering contribution: distinguishing correlates from causes

One major contribution of this review is the introduction of an evidence-tiering system to evaluate support for each hypothesized cause of bullying. Past reviews have typically enumerated risk factors without discriminating the rigor of evidence behind them, often giving equal weight to findings from disparate methodologies (Clayton R Cook et al., 2010; Hong & Espelage, 2012; Zych et al., 2020). In contrast, we systematically categorized studies from Tier 1 (cross-sectional associations) through Tier 5 (randomized experiments), adapting an epidemiological lens to the social science context. This hierarchical evaluation is theoretically significant because it allows us to discern which relationships are consistently observed and which are actually likely to be causal. By distinguishing mere correlates from well-substantiated causal factors, we provide a clearer map of bullying’s etiology that can guide both theory and intervention (Hall, 2017; Juvonen & Graham, 2014; Swearer & Hymel, 2015). Notably, our synthesis revealed that for many commonly cited risk factors, such as low empathy, academic problems, or peer delinquency, the literature is rich in correlations (Tier 1) but scant in robust longitudinal or experimental tests (Tier 3-5) that could confirm causation. The dearth of high-tier evidence in certain areas highlights an important knowledge gap: much of what we “know” about bullying’s causes rests on modest inferential grounds (Maxwell & Cole, 2007).

By making these evidentiary gaps explicit, our review challenges theorists to refine causal claims and encourages researchers to apply stronger designs to test those claims (Chalmers & Glasziou, 2009; Durlak & DuPre, 2008; Munafò et al., 2017). In essence, this evidence-tiered approach elevates the theoretical discourse around bullying by grounding it in the principles of causal inference, ensuring that theoretical propositions about why bullying occurs are evaluated in light of how rigorously they have been examined.

5.3 Global generalizability and knowledge inequity

In taking a global perspective, our review also critically examines whether the dominant theories and findings, largely derived from Western contexts, generalize worldwide (Henrich et al., 2010). Bullying research to date has been disproportionately conducted in North America, Europe, and Australia, with far less representation from low- and middle-income countries (Hong & Espelage, 2012; Menesini & Salmivalli, 2017; Smith, 2016). This imbalance creates a form of “knowledge inequity” whereby certain cultural assumptions may be treated as universal in the literature without sufficient testing elsewhere (Arnett, 2008; Norenzayan & Heine, 2005). If researchers impose one interpretive lens globally, they risk overlooking context-specific risk dynamics. Our findings suggest that some risk factors, such as weak school bonding or peer deviance, appear robust across many countries studied, whereas others show variability (Biswas et al., 2020; Koyanagi et al., 2019). For instance, the prevalence of bullying related to bias, e.g. racial or ethnic harassment, might be higher in societies with salient intergroup conflicts, whereas in other contexts relational or cyber forms may predominate (Russell et al., 2012; Smith, 2016). We also note that reported bullying rates differ widely by country, not only because of true behavioral differences but also due to measurement and cultural norms around reporting (Chester et al., 2015; Craig et al., 2009; Lee & Larson, 2000). These observations underscore that theoretical models of bullying should be applied with cultural sensitivity. An explanatory factor may hold in one context but need modification in another, an insight echoed in cross-national comparative studies (Fleming & Jacobsen, 2010; Zych et al., 2020).

Therefore, one practical outcome of a global lens is the recognition that anti-bullying strategies must be tailored: interventions successful in Finland or the U.S. might require adaptation to local values, resources, and bullying profiles in, say, South Asia or sub-Saharan Africa (Bowes et al., 2009; Kaufman et al., 2018; Sivaraman et al., 2019). In line with this, there is a pressing need for more longitudinal and high-quality studies from under-represented regions to validate and enrich existing theories (Biswas et al., 2020; Menesini & Salmivalli, 2017; Smith, 2016). Expanding the geographic scope of research is not just an issue of equity but will refine the theoretical understanding of bullying by revealing which causal factors are truly universal and which are culture-bound.

5.4 Implications for prevention and limitations for research

The insights from this review carry several concrete implications for prevention, policy, and future research. First, identifying the most credible causal risk factors provides a roadmap for prevention programming (Cho et al., 2019; Clayton R Cook et al., 2010; ten Bokkel et al., 2023). Schools and communities should focus on evidence-based levers, such as improving peer climates, strengthening teacher-student relationships, and building self-regulation skills, that our synthesis suggests are causally linked to reductions in bullying (Bradshaw et al., 2015; Gaffney et al., 2021; Salmivalli et al., 2011; Ttofi & Farrington, 2012). For example, given robust evidence that peer influences contribute to bullying, interventions like peer-led anti-bullying campaigns or bystander empowerment programs are theoretically well-founded strategies to disrupt group reinforcement of aggression (Paluck et al., 2016; Polanin et al., 2022; Salmivalli et al., 1996). Similarly, because weak social bonds and poor school climate emerge as risk factors, whole-school interventions that foster inclusion, respect, and connectedness, such as KiVa or the Olweus Bullying Prevention Program, have a strong conceptual basis and have demonstrated effectiveness in multiple trials (Kärnä et al., 2013; Kärnä et al., 2011; Nocentini & Menesini, 2016). Our review’s emphasis on causal drivers also helps policymakers prioritize upstream solutions over reactive measures (Bradshaw et al., 2015; Hensums et al., 2023). It suggests that policies addressing root causes, like reducing school overcrowding to improve monitoring, mandating social-emotional curricula, or implementing anti-bullying laws with enforcement mechanisms, could yield population-level reductions in bullying (Durlak et al., 2011; Hatzenbuehler et al., 2015). More broadly, viewing bullying through a causal lens has a vital societal impact: it counters fatalistic attitudes that peer aggression is an inescapable part of childhood and instead promotes the understanding that reducing bullying is an achievable goal through deliberate changes in environments and behaviors (Gaffney et al., 2021; Salmivalli et al., 2011). By highlighting factors beyond the individual, such as school climate, policies, and peer group norms, this perspective helps shift the narrative away from stigmatizing victims or dismissing bullying as mere “kids will be kids” conflict, and toward a focus on collective responsibility and systemic solutions (Cornell & Limber, 2015; Paluck et al., 2016). Such a reframing can reduce victim-blaming and mobilize educators, parents, and students alike to collaborate in creating safer social environments.

Second, our analysis draws attention to the limitations of current research and the opportunities for advancing knowledge. One key limitation is the relative scarcity of high-quality longitudinal, quasi-experimental, and experimental studies targeting bullying causes. While there are hundreds of cross-sectional studies each year, far fewer studies employ designs that support strong causal inference (Farrington et al., 2017; Kljakovic & Hunt, 2016; Zych et al., 2020). This imbalance means that some widely accepted “risk factors” might not truly be causal, or might have much smaller effects than presumed once confounding is accounted for. For the field to progress, researchers should invest in methodologies such as: long-term prospective cohorts that can establish temporal sequences; within-family or within-school comparisons (e.g., twin studies, or comparing cohorts before and after an intervention) to control for unmeasured confounds; natural experiments leveraging policy changes or school reforms; and randomized controlled trials of interventions that test specific causal hypotheses (Craig et al., 2017; D’Onofrio et al., 2013; Stuart, 2010). Encouragingly, some recent efforts are moving in this direction, evaluations of anti-bullying programs double as tests of causal theories, e.g., if boosting empathy or improving supervision reduces bullying, it substantiates those factors’ causal roles (Gaffney et al., 2021; Kärnä et al., 2011; Salmivalli et al., 2011). We recommend that future research not only continue to assess “what correlates with bullying” but increasingly ask “what happens to bullying rates if we change X?”, whether X is a classroom management practice, a parenting intervention, or a peer norm campaign (Angrist & Pischke, 2010; Espelage et al., 2014; Hernán, 2018). Additionally, standardizing measurement of bullying across studies and countries would address a current limitation where divergent definitions and instruments hinder direct comparison and pooling of data (Liu et al., 2025; Vaillancourt et al., 2008; Volk et al., 2014; Zhao et al., 2022). Developing a core set of validated bullying measures for international use, for instance, a common module of bullying items added to youth surveys globally, could greatly enhance our ability to compare prevalence and test theories cross-culturally (Chester et al., 2015; Jinxiao Ginnie, 2024; X. Li et al., 2025; Smith et al., 2002). Such harmonization, combined with data sharing and collaborative multi-site studies, represents a way forward to overcome the limitations of siloed, incomparable studies that currently dominate the field.

Finally, it is important to acknowledge the limitations of this review itself and outline directions for future inquiry. While we attempted to be comprehensive, our review may be constrained by publication bias in the literature, studies finding significant predictors of bullying are more likely to be published, potentially overestimating the importance of certain factors (Farrington et al., 2017). We focused on adolescent populations; thus, we may have excluded relevant insights from younger age groups, or unpublished studies that could nuance our conclusions. Additionally, our evidence-tier categorization, while useful for clarity, is an imperfect tool: study quality varies within each tier, and even a Tier 4 study, e.g., a quasi-experiment, can yield misleading results if not well executed. We treated the presence of higher-tier evidence as indicative of stronger causal support, but readers should note that causality can only be inferred, not proven with absolute certainty, especially in social contexts (Greenland, 2017; Hernán, 2018). Another limitation is that our narrative synthesis could not quantitatively weight the effects of each factor; a formal meta-analysis was beyond our scope, given the heterogeneity of measures and designs, but future meta-analytic work could build on our categorization to estimate the magnitude of causal effects where possible (Gaffney et al., 2021; Gini et al., 2014; Ttofi & Farrington, 2011, 2012; Zhao et al., 2010). We also recognize that bullying is an evolving phenomenon, especially with technology creating new modalities of cyberbullying, and the relative importance of different causes may shift over time or differ for cyber versus traditional bullying (Kowalski et al., 2014; Munafò et al., 2017; Olweus, 2013). There is a need for continuous updating of evidence as new social platforms, educational policies, and cultural trends emerge that could influence bullying dynamics.

5.5 Promotion for global reciprocal, co-produced cross-regional research

This synthesis forces a reckoning with the power relations that quietly organize “global” bullying research (Arnett, 2008; Henrich et al., 2010). When Western canonical criminological frameworks are extended into Asia and the Global South as if they were default universals, local settings are too often positioned as mere testing grounds and local concepts as noise to be translated away (Carrington et al., 2016; Liu, 2009, 2021). In that moment, the field risks enacting an epistemic analogue of the very phenomenon it studies: a form of scholarly bullying in which prestige defines the terms of debate, peripheral voices are pressured to conform, and difference is treated as deficit (O'Connell et al., 1999). It is an argument for reciprocity and humility: theories should earn their universality by surviving contact with distinct histories, institutions, moral orders, and relational ecologies.

Empirically, we need designs that move beyond first-order correlations and can isolate mechanisms across contexts, including natural experiments, network-based interventions, and longitudinal models that treat interdependence as a feature (Faris & Felmlee, 2011; Paluck et al., 2016; Snijders et al., 2010). Epistemically, we need a more plural architecture of explanation in which Asia and the Global South are sources of it: places where new constructs can be named, new causal pathways can be specified, and old assumptions can be revised (Carrington et al., 2016; Cornell & Limber, 2015; Liu, 2009, 2021; O'Connell et al., 1999). A truly global science of bullying will be built by co-producing knowledge across unequal academic geographies, treating translation as theory-building, and designing comparative work that protects conceptual dignity as carefully as it pursues measurement equivalence (Faris & Felmlee, 2011; X. Li et al., 2025; Liu & Li, 2024; Volk et al., 2014). If bullying is fundamentally about the abuse of asymmetry, then our closing imperative is ethical as well as scientific: to ensure that a review about bullying does not reproduce, at the level of knowledge, the silencing and hierarchy that bullying inflicts in everyday life.

6. A forward agenda: designs and pitfalls

6.1 Credible claims about causality so far

Most of what is “known” about why bullying occurs still rests on correlational evidence, which cannot by itself identify causes. Cross-sectional associations (Tier 1) are useful for mapping who is involved and what conditions co-occur with bullying, but they are vulnerable to confounding, measurement variation, and reverse causality. Bidirectional processes are common: victimization can worsen mental health, while pre-existing distress can also increase later victimization, producing feedback loops rather than one-way pathways (Karyotaki et al., 2021; Reijntjes et al., 2011).

Prospective longitudinal studies (Tier 2) improve on cross-sectional work by establishing temporal order, and meta-analytic syntheses show that some factors repeatedly precede later bullying outcomes (Kljakovic & Hunt, 2016; Zych et al., 2020). Yet temporality is not causality (Robins et al., 2000). Many “predictors” may reflect unmeasured, time-varying confounders, and replication remains uneven because many determinants have been examined longitudinally only once or twice (Cole & Hernán, 2008). School-level features such as climate can forecast later bullying, but even strong prospective correlations rarely distinguish whether a factor is causal, a proxy for other processes, or part of a reciprocal dynamic (Hamaker et al., 2015; Low & Van Ryzin, 2014). Accordingly, most etiological theories still draw heavily on Tier 1-2 evidence: informative for sequencing and prioritization, but insufficient for confident causal claims.

The clearest causal leverage comes from the relatively small Tier 3-5 literature, including within-unit designs (fixed effects; sibling/twin comparisons), quasi-experiments, and randomized trials. Within-person analyses can rule out stable individual differences, strengthening inference when changes in a putative determinant track changes in bullying (Cho & Lee, 2018; Freese & Kevern, 2013). Genetically informed designs also suggest non-trivial environmental contributions to bullying above and beyond heritable propensity (Ball et al., 2008; Brendgen et al., 2013). Quasi-experimental policy variation (e.g., staggered legal or program roll-outs) and whole-school trials provide stronger tests of mechanisms; the latter consistently imply that modifying peer norms, supervision, and school climate can reduce bullying (Kärnä et al., 2011; Paluck et al., 2016; Ttofi & Farrington, 2011). By contrast, several popular explanations still lack credible causal support: community constructs such as collective efficacy may be endogenous to local violence and disorder (Hipp & Wickes, 2017; Sampson et al., 1999); and adversity/ACE measures are often treated as causal “dose” scores despite substantial ambiguity about timing, specificity, and confounding (Jaen et al., 2023; Lacey & Minnis, 2020).

In short, causal evidence is uneven: peer- and school-system mechanisms and self-regulation have comparatively stronger multi-tier support, while other domains remain largely correlational. The next step is investing in designs that can move key hypotheses into Tier 3-5 evidence and clarify when effects generalize across contexts.

6.2 Methodological directions and cautionary notes

Moving the field toward stronger evidence will require pairing better designs with better analysis. Modern causal inference tools, such as double machine learning and causal forests, can help adjust for high-dimensional confounding and explore heterogeneity, but only when embedded in credible identification strategies and paired with clear estimands (Athey & Imbens, 2017; Hernán, 2018; Rubin, 1974; Wager & Athey, 2018). DML can “de-bias” treatment-effect estimation after flexible nuisance modeling, making it attractive in longitudinal panels or quasi-experimental settings with many covariates (Bang & Robins, 2005; Chernozhukov et al., 2018). Causal forests can estimate conditional average treatment effects, allowing researchers to test whether effects differ by baseline risk, school climate, or other moderators without manually specifying many interactions (Athey & Imbens, 2017; Wager & Athey, 2018). Used appropriately, these tools maximize learning from complex school and survey data; used carelessly, they risk laundering weak design into overconfident causal language. Researchers should also explicitly consider interference and spillovers, especially in school settings where one student’s treatment or victimization can affect others. When interference is plausible, cluster designs and models that accommodate network or classroom spillovers should be the default (Basse & Feller, 2018; Hudgens & Halloran, 2008; Sinclair et al., 2012).

Conceptually, researchers should avoid shortcuts that blur interpretation. Summed ACE indices are often interpreted as mechanistic causal scales, yet they combine heterogeneous experiences with different causal pathways and can invite deterministic or stigmatizing narratives (Anda et al., 2020; V. J. Felitti et al., 1998). More informative work disaggregates adversities, specifies timing, and tests plausible mechanisms rather than relying on an all-purpose score (Jaen et al., 2023; Lacey & Minnis, 2020). Likewise, correlations between collective efficacy and bullying do not establish that “building cohesion” will reduce bullying; reciprocal dynamics and endogeneity are plausible until tested with longitudinal sequencing or stronger designs (Hipp & Wickes, 2017; Sampson et al., 1997). Precision about constructs, what exactly is being changed, by whom, and over what time frame, should precede causal estimation.

Finally, progress depends on shifting from post hoc theory-fitting to genuine theory-testing. Bullying theories should generate risky, a priori predictions that can be falsified and compared against alternatives, ideally in designs with clear counterfactual leverage (Kerr, 1998; Munafò et al., 2017; Nosek et al., 2018). This stance aligns evidence generation with mechanism: studies can specify mediators and timing, test competing pathways, and report null or contradictory results as informative (Altman & Bland, 1995). In practice, this means designing studies that can adjudicate between mechanisms, through multi-arm trials, natural experiments, or longitudinal mediation tests, and pairing them with transparency practices (e.g., pre-specified hypotheses and open materials) that reduce hindsight bias. A cumulative science of bullying causes will therefore look less like a catalogue of correlates and more like an iterative program of design-based tests that refine theory while informing practical prevention.