Bullying roles overlap in teens, and bystander behavior may hold the key

Adolescents who bully, who are bullied, and who stand by and watch are often treated as three separate populations. A new study argues that this tidy division hides a messier reality: within the same teenager, all three forms of involvement can cluster together far more often than chance would predict. Using data from nearly 8,000 Chinese secondary school students, researchers mapped the specific behaviors that link victimization, perpetration, and bystander reactions into a single interconnected system, and then ran computer simulations to identify which behaviors might offer the biggest leverage for future interventions.

The study, published in Frontiers in Psychiatry, analyzed survey responses from 7,913 students collected between April and July 2018 in a province of south-central China. The original sample included 9,156 invited students, of whom 8,439 completed the questionnaire, a response rate of 92.2 percent; after screening out incomplete or logically inconsistent responses, 7,913 valid questionnaires remained. The group included 3,766 boys and 4,147 girls, roughly evenly split between rural and urban schools and between junior and senior high school. Because of privacy policies, the survey did not collect students’ ages, so no age statistics are reported.

The researchers first asked a simple question: how often do victimization, perpetration, and bystander responses occur together in the same student, and does that exceed what would be expected if the three dimensions were independent? Each dimension was coded as present or absent, yielding eight possible patterns. The most common pattern, by far, was bystander responses alone, reported by 66.17 percent of students. But the striking result concerned the full overlap. All three dimensions co-occurred in 3.16 percent of students, roughly four times the frequency expected under independence, with an observed-to-expected ratio of 4.08 and a 95 percent confidence interval of 3.75 to 4.43. Even the victimization-perpetration overlap alone, though rare at 0.40 percent of the sample, occurred about two and a half times more often than chance would suggest.

To move beyond broad categories, the team turned to network analysis. They used the Ising model, a statistical framework designed for binary variables, to estimate conditional associations among 26 individual behaviors: nine victimization items, nine perpetration items, and eight bystander responses drawn from the revised Olweus Bully/Victim Questionnaire and the Student Bystander Behavior Scale. Each item had been dichotomized according to the established Olweus frequency criterion, marking behaviors that occurred two or three times a month or more often. The resulting network contained 75 nonzero edges, 73 of them positive, and 14 edges connecting nodes across the three behavioral communities.

The strongest cross-community connection was between verbal teasing on both sides: teasing others was strongly linked to being teased oneself. Rumor spreading also stood out, with perpetration of rumor spreading showing the highest strength centrality in the network, followed by victimization through rumor spreading and the pro-bully bystander behavior of joining in bullying. Pro-bully responses such as joining in, assisting the bully, and cheering or ridiculing the victim showed positive conditional associations with perpetration, consistent with the idea that assisting and reinforcing bullying share behavioral content with bullying itself. In contrast, the reverse-coded defender responses, failure to intervene and failure to report to adults, showed limited direct connections with perpetration, and two cyberbullying perpetration items with very low prevalence emerged as isolated nodes in the binary network, though a supplementary analysis using the original ordinal responses suggested this isolation was sensitive to the coding choice rather than evidence of true behavioral independence.

The study’s most technically novel step came next. Centrality measures identify well-connected nodes, but they do not capture what happens if a behavior actually changes. For that, the researchers applied Node Identification and Response Analysis, or NIRA, a simulation-based algorithm that systematically shifts the threshold parameter of each node in the fitted Ising model, in this case by two standard deviations in the direction that increases the modeled probability of activation, and then generates new response samples to see how overall network activation changes. For each of the 26 nodes, the team compared a baseline sample with a perturbed sample using independent-samples t tests, with p values adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate procedure.

In the primary analysis, the perturbed node itself was excluded from the outcome, so the measured change reflected the remaining 25-item aggregate score. Five bystander-response nodes produced statistically detectable reductions after FDR correction, and all five were bystander behaviors. The largest projected change came from the failure-to-intervene node, where the aggregate score dropped from 2.11 to 1.57, a difference of minus 0.54 points. The failure-to-report node followed with a change of minus 0.43, then leaving the situation at minus 0.29, pretending not to see at minus 0.24, and watch passively at minus 0.15. Notably, these low-centrality nodes, which sat at the periphery of the estimated network, produced larger simulated effects than the highly connected perpetration and victimization behaviors, none of which reached significance in this analysis.

Yet the authors are careful about what these simulations do and do not show. In role-specific subscale analyses, the projected changes from perturbing bystander nodes were largely confined to the bystander-response subscale itself. Perturbing the failure-to-intervene node, for example, produced a projected difference of minus 0.53 points in the remaining bystander items but changed the victimization and perpetration subscales by essentially nothing. Simulated mitigation of individual perpetration nodes did not significantly change the remaining perpetration items at all. The authors conclude that bystander-response nodes may represent candidate leverage points for changing related bystander behaviors, rather than established targets for reducing victimization or perpetration.

The limitations are substantial and the authors state them plainly. The data are cross-sectional, so the network reflects conditional associations, not dynamic processes or causal directions, and the findings should not be read as evidence that one role transitions into another. The study relied on student self-reports, which are vulnerable to recall bias, social desirability, and underreporting of socially disapproved behavior. NIRA identifies nodes that produce large simulated effects under a specified network structure and perturbation rule; it does not establish real-world intervention targets. The single-level Ising model also did not explicitly partition the classroom- and school-level clustering that variance-component analysis had quantified, and the candidate behaviors were not externally replicated or tested in longitudinal or experimental designs.

What the study offers, then, is a set of testable hypotheses rather than conclusions. The authors suggest that future school-based intervention research might consider defender bystander responses, such as safe intervention and reporting to adults, as candidate targets, and that anti-bullying programs might address the clustering of multiple forms of involvement rather than treating perpetrators, victims, and bystanders as mutually exclusive categories. They also flag rumor spreading, given its central position in the network, as a less overt form of aggression that prevention efforts might monitor. Whether promoting active bystander intervention actually reduces bullying involvement, and whether such effects endure, remains to be tested in class-level randomized or quasi-experimental studies, which would also need to weigh potential risks such as peer retaliation against students who report bullying.

Subject of Research: A cross-sectional network analysis of co-occurring bullying victimization, perpetration, and bystander responses among Chinese adolescents

Article Title: The co-occurrence of victimization, perpetration, and bystander responses in adolescent school bullying: a network analysis with model-based perturbation simulations

Article References: Qiu, X., Liu, X., Miao, Y., Jiang, P., & Gao, R. (2026). The co-occurrence of victimization, perpetration, and bystander responses in adolescent school bullying: a network analysis with model-based perturbation simulations. Frontiers in Psychiatry, 17, Article 1994414. https://doi.org/10.3389/fpsyt.2026.1994414

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1994414

Keywords: school bullying, adolescents, network analysis, victimization, perpetration, bystander behavior, NIRA, Ising model, co-occurrence, cross-sectional study, China, simulation