Past research on cyberbullying among adolescents largely consisted of surveys that measure the incidence of online aggression and identify preferred cyberbullying modes, such as social media sites. That work stopped short of closely examining cyberbullying message content and probing the peer network dynamics around those messages. In that respect, it largely lacked the kind of evidence-based information that could open more windows into the nature and impact of cyberbullying and help point to solutions for peers, parents, schools, and law enforcement.
A recently completed multimethod study of cyberbullying has combined analysis of actual message content with measures of frequency, peer group context, and the correlation between face-to-face and electronic bullying.
The researchers worked with the following standard definition of bullying: “aggressive behavior that is repetitive and results in a power differential.”
The findings and methods of the research team — led by Marizen Ramirez and Anthony Paik and supported by an NIJ grant — offer new insights for school leaders and others charged with identifying and responding to bullying, according to a report from the researchers. These insights, which may help identify and inform responses to bullying, include:
- An understanding that assessing exposure to cyberbullying may require multiple measures — self-reports produced a different picture of exposure than analyses using smartphone data.
- Identification of the kind of message content that constitutes prohibited bullying — a potential benefit to many schools struggling to define, identify, and stop prohibited bullying conduct.
- A finding of a strong association between face-to-face bullying and online bullying — a result that can inform school safety policies to help ensure that investigations of bullying in schools extend to electronic communications, the researchers said.
- A finding that messages with negative, bullying content tended to be clustered among key individuals — a new insight suggesting that a small number of individuals often play vital roles in spreading negative messages across social networks.
- Progress toward development of a new instrument with potential for use by administrators to cull thousands of messages and code aggressive content — a resource that has practical implications for bullying interventions and “for the criminal justice response and management of ‘big data’” though further work is needed, the research report said.
- A finding that students who are bullied tend to be well integrated in, but not central to, school peer networks — a new insight suggesting an opportunity to focus adolescent bullying studies on students who are close to peer groups, rather than those who are socially more isolated.
A notable, if not surprising, result of the study was the sheer volume of electronic messages generated by the sampled sixth through eighth graders. The researchers noted that the studied teens engaged in a huge volume of cyber communications — about 2,000 messages per participant over one school semester.
Growing concerns over cyberaggression track the rapid evolution of electronic messaging by adolescents. The report cited a study stating that, by 2013, 78% of teens used cellphones, half of which were smartphones. 
No less than 95% of U.S. teens ages 12 to 17 were online using email, social media sites, or texts, according to another 2013 study cited in the research report. 
Notwithstanding the proliferation of electronic communications among teens, “Cyberspace as an avenue for receiving or delivering acts of aggression is not well understood,” the researchers reported. A limitation of the extant literature, the report said, was the absence of gathered data on the cyberbullying content of electronic communications and the peer context of those communications. The research team’s intent was to unearth the foundational cyberbullying content and context needed to establish evidence-based interventions.
The research team’s objectives were to: (1) classify the contents of cyberbullying messages and measure the frequency of various themes; (2) estimate associations between cyberbullying and offline bullying; and (3) examine associations between cyberbullying and participants’ positions in social networks, as well as network composition.
The researchers used both surveys and electronic capture of smartphone data to gather cyberbullying data from a total of 164 students in grades six through eight at two Iowa middle schools, with parental agreement. The researchers conducted one survey at the start of the spring semester and another at the semester’s end. Electronic data were captured from smartphones used by 77 of the 164 study participants. The smartphones were set up to collect all incoming and outgoing text messages. For Android users, the research application collected all texts as well as Facebook and Twitter messages. For iPhone users, an application captured Facebook and Twitter posts, but no text messages — a technical limitation of the study. Researchers surveyed smartphone participants weekly on their bullying experiences as victims, individuals who bully, or witnesses.
The study employed a number of rigorous research methods, approaching key inquiries from multiple angles. Those methods included self-reports, electronic capture of information, electronic weekly queries, social network analysis, and mixed methods — a research approach combining quantitative and qualitative data analysis — as well as machine learning. (Machine learning is a form of artificial intelligence that uses algorithms to build mathematical models from sample data and then make predictions or decisions about the full dataset. In machine learning, the machine “learns” from the sample, or training, data. The program’s performance improves with experience.)
The survey component of the study had two phases — an initial survey in December 2014 and a follow-up survey in May 2015. Participants were asked for their demographic information as well as experiences such as daily activities, health, academics, and attachment to parents. They were also questioned about their electronic communications, experiences of aggressive behavior, delinquency, and drug and alcohol use.
Smartphone messages were collected and measured in two ways:
- Weekly queries — Researchers queried users each week to collect messages. When a participant reported being a victim or witness of cyberaggression, researchers also drew that participant’s messages from the preceding week and the following week into the case sample. The team created a control set of messages gathered from participants who reported experiencing no cyberaggression. Two researchers then independently coded all of the sampled aggressive messages, using content analyses to measure aggression, reasons for any aggression (i.e., appearance, race, ethnicity, religion, gender, and sexual orientation), and role of the participant (e.g., victim, individual who bullies, or bystander). The research team also queried the smartphone user group each week about bullying experiences and incorporated their responses into the research data.
- Paired messages analyzed through machine learning — A total of 110,040 messages, between pairs of participants who had exchanged at least 10 smartphone messages, were analyzed for aggressive content.
Cyberaggression Comprised a Relatively Low Percentage of All Electronic Messages
Overall, the research team found that a relatively small number of electronic messages between teens reflected aggressive behaviors. These counts varied, however, depending on how data were collected:
- In the self-report surveys, 5% of students had experienced cyberbullying in the preceding two months.
- In the content analysis of a sample of 35,566 messages subjected to qualitative coding, including coding of aggressive forms of communication, 8.1% were coded as aggressive in content.
- In the machine-learning analysis of messages between pairs of participants — a total of 110,040 messages — 2% were coded as aggressive in content.
For the messages between participant pairs, machine-learning techniques allowed instant isolation of those messages containing term clusters identified as being associated with aggressive behavior. An algorithm classified documents on the basis of measures including a common statistical measure known as “term frequency – inverse document frequency” that identifies and ranks documents in a set according to both the presence of search terms in a given document and the relative absence of those same terms in all other documents. In that way, the technique can flag for investigators those documents reflecting aggressive expressions or conduct.
Fighting Was the Most Common Topic of Aggressive Messages
Within the “aggressive” 2% of messages between pairs of participants, the most common topic was fighting, at 36.4%. Aggressive bullying messages were less likely to be related to personality traits, sexual activity, harassment, jealousy, and appearance.
Victims of Face-to-Face Bullying More Likely To Be Cyberbullying Victims
Of the survey sample of 164 students, 21 (15.3%) reported being victims of face-to-face bullying, while 5.1% reported in the initial survey that they had been victims of cyberaggression, compared with 5.4% at follow-up. Surveyed students who reported face-to-face bullying had 7.31 times the odds of subsequently being cyberbullying victims than those students who did not experience face-to-face bullying.
In the sample of students who used smartphones, researchers found a cyberaggression witness rate of 14.2 reports per 100 student-weeks, and a victimization rate of 5.4 reports per 100 student-weeks. Those in the smartphone sample who were bullied face-to-face had 3.7 times the rate of cyberbullying victimization of those who did not report being bullied face-to-face.
Negative Sentiment Clustered Among Participants
The research focused on both the linguistic and relationship contexts of cyberbullying. Electronic messages with negative sentiment were clustered among participants, indicating that a few individuals are instrumental in moving negative sentiment across the larger contact group. That finding is consistent with an understanding that cyberbullying is more likely to be targeted at certain individuals as opposed to being a product of mutual conflict, the report said. The research also brought to light the role of gossip as a major aspect of cyberbullying: Negative sentiment was strongly associated with discussion of third parties.
A Potential New Technology for Coding Aggressive Cyber Communications
The team developed new software for coding and identifying aggressive messaging found in large volumes of texts and social media traffic, a process faster and more efficient than manual content review by research team members. The researchers said that the software has potential implications going forward for bullying interventions, criminal justice activity, and management of big data. However, further research on and attention to data-sharing concerns are needed.
Peer Networks Present as Possible Intervention Points
Social network analyses indicated that bullying victims tended to be close to the center of social networks but not central themselves, suggesting that peer networks can be an intervention mechanism for promoting prosocial — as distinct from antisocial — norms.
Though the sample size for the study was small, it provided valuable information on a topic of great interest that has seen only limited research. The research team’s new window into what cyberbullying message content looks like could raise awareness of the kinds of messages that are harmful — a subject that has challenged educators and others charged with enforcing bans on bullying. This research helps lay the groundwork for developing anti-cyberbullying interventions, including actions by bystanders who witness bullying.
About This Article
The research described in this article was funded by NIJ award 2013-IJ-CX-0030, awarded to the University of Iowa. This article is based on the grantee final report “Contents and Contexts of Cyberbullying: An Epidemiologic Study Using Electronic Detection and Social Network Analysis,” Marizen Ramirez and Anthony Paik, Principal Investigators.