Subjects and procedure
This study was part of the Finnish Nationwide 1981 Birth Cohort Study (Figure S1). The Joint Commission on Ethics of Turku University and Turku University Central Hospital approved the research plan. Participation was voluntary and based on informed consent from the children’s parents. The basic population was all 60,007 Finnish-speaking children born in Finland in 1981. At this time, 93.5% of people who lived in Finland, were Finnish-speaking [16]. A random sample of about 10% of the basic population was drawn, including rural, suburban and urban communities. The sample comprised 6017 children and 5813 (96.6%) took part. The baseline assessment was conducted in 1989, and information was obtained from the children, their parents and teachers [17]. Parents received information and questionnaires from the school via their child and returned their completed questionnaires to the teachers in a sealed envelope. The children filled in their questionnaires in the classroom and the teachers completed their questionnaires after parental consent. The teachers returned the study material to the research group.
Information on bullying
Questions on bullying perpetration and victimization were included in the questionnaires [10,11,12, 18]. Children were asked to recall any bullying incidents during the past two weeks. Bullying was assessed by giving children three alternatives to choose from: “I bully other children almost every day”, “I bully sometimes” and “I do not usually bully”. Correspondingly, the alternatives to assess victimization were: “Other children bully me almost every day”, “Other children bully me sometimes” and “Other children do not usually bully me.” Parents and teachers were asked if the child had been a bully or a victim over the last 12 months and the responses were certainly applies, applies somewhat and does not apply. In the statistical analyses, the options indicated frequent involvement, some involvement or no involvement in bullying, respectively.
The analyses used pooled information from the children, parents and teachers. The perpetrators and victims were each split into three groups: no involvement, some involvement and frequent involvement, based on the highest rating of bullying by any informant. We also analyzed whether the associations between bullying and any violent offenses were different for bullies and bully-victims. Perpetrating bullying sometimes or frequently was recorded as being a bully or a bully-victim, and this depended on whether the subjects had also been victims of bullying. We formed a bullying variable with three categories for this additional analysis. The categories were those who had not bullied at all, but may have been victims of bullying, those who had just bullied and those who had been both bullies and victims of bullying.
Covariates
The parents’ questionnaire included information on the child’s sex, parental education and family structure. The teachers’ questionnaire included information on the child’s psychopathology. These were used as covariates in the analyses. Parental education was based on whether or not one or both of them had completed upper secondary school education. The family structure options were living with two biological parents, one biological parent, one biological parent and a step-parent, foster parents, adoptive parents or some other family model. These were dichotomized for the statistical analyses into whether or not the child lived with two biological parents. Psychopathology was measured with the Rutter Teacher Questionnaire. This was reported to be more valid than the Rutter Parent Questionnaire when screening for psychiatric disturbances in the baseline sample of the study cohort [19] and was shown to have satisfactory reliability [20]. The Rutter Teacher Questionnaire includes an item on bullying perpetration and this item was removed for the analyses. The psychopathology scale was then dichotomized, with nine or more points indicating psychiatric problems. This was based on previous epidemiological studies [19, 20]. Because the bullying item was removed, we also conducted the analyses using eight points as the cut-off point to validate the results.
Information on violent offenses
Information on violent offenses was obtained from the Finnish National Police Register, an electronic database that includes all cases where the police have suspected a named individual of an offense. The National Police Board provided permission. The registration threshold is low and if a person is suspected of multiple offenses, they all are registered. The data are archived after the window of time for prosecution has elapsed. Both the police register and the archive were examined to get comprehensive data that covered the follow-up period. The register data were linked with the study subjects by the personal identification code given to all Finnish citizens. This was approved by the Data Protection Ombudsman.
The age of criminal responsibility in Finland is 15 and that is why the observation period began at this age. The data were collected on 3 May 2012. The register includes the actual dates when the offenses were committed and this allowed us to monitor the follow-up period on a daily basis. Of the 5813 children who took part in the baseline study, seven had died and four had emigrated before they were 15 and the identification codes needed to link individuals to the police register were missing in 397 cases. We were able to check 5405 subjects (50.3% male) and this showed that 515 (9.5%) were registered for violent offenses, 36 had died and 95 had emigrated during the follow-up period. These 5405 subjects represented 89.8% of the original sample and 93.0% of those who participated in 1989. At the time of the data collection, 65.5% were aged 30 and 34.5% were 31.
The data on the crime register data included the crime category, which enabled us to identify violent offenses. Violent crime was defined as overt aggressive behavior toward another person and divided into minor and severe violent offenses (Table S1). The categorization was based on Finland’s Criminal Code. By dividing offenses into minor and severe violent offenses, we aimed to distinguish between the most severe forms of violent offenses from a legal point of view. For example, homicide was included in severe violent offenses. Attempted crimes were included in the respective groups.
Statistical methods
The descriptive statistics of the study sample and violent offenses were calculated. The correlations between bullying perpetration and victimization, and the agreement between the informants, were calculated using dichotomic variables that indicated whether or not the child had been a bully or a victim. These were tested for the total sample, by including both males and females. We calculated the correlations using pooled information from the children, parents and teachers. Cohen’s kappa coefficient (κ) was used to calculate the level of agreement about bullying and victimization between the parents and children, the parents and teachers and the children and teachers. Attrition analysis was carried out for the sample that included both males and females. The characteristics of the study sample and the 1989 attrition group were compared. This included bullying perpetration and victimization, sex, whether or not the parents had completed upper secondary school education, the family structure and child psychopathology. The bullying perpetration and victimization variables were divided into three categories, as described above, and Pearson’s chi-square test was used to carry out the attrition analysis. The other variables were dichotomized and based on the coding described above. Fisher’s exact test was used to conduct these analyses.
Statistical models were created to predict any violent offenses and severe violent offenses. The explanatory variables were being a childhood bully or a victim. There were three categories that indicated involvement in bullying. These were never being a bully or a victim or being a bully or a victim sometimes or frequently. The separate sex × bully and sex × victim interactions for violent offenses were not significant, but we conducted separate sex-based analyses due to the differences in both bullying involvement [13] and violent criminality [14, 15]. Bullying was controlled for victimization and vice-versa. These variables were also divided into three categories that indicated involvement in bullying. These were never being a bully or a victim or being a bully or a victim sometimes or frequently. Parental education level, family structure and child psychopathology were also used as covariates. Hazard ratios (HRs) and 95% CIs were estimated using Cox regression analyses [21]. Survival time was defined as the amount of time that had elapsed from 15 years of age to the first event. The level of accuracy that was used to define time was 1 day, because the exact dates of the offenses were registered. The first event was the first violent offense of any severity when we analyzed any violent offenses. The first severe violent offense was used when we analyzed severe violent offenses. The subjects were censored at time of death or moving abroad or at the end of the study period if they had not been registered for offenses.
We carried out sensitivity analyses to separately estimate the associations between bullying perpetration or victimization at 8 years of age, as reported by the children, their parents and teachers, and any violent offenses by the age of 31. We used the dichotomic bullying and victimization variables to indicate whether or not the child had been a bully or a victim. The outcome variables fell into two categories and these were not committing any violent offenses, which was the reference category, or committing any violent offenses. Single predictor binary logistic regression models were carried out to estimate the odds ratios (ORs) and 95% confidence intervals (CIs).
We also carried out some additional analyses. First, the additional analyses assessed whether the OR for violent offenses in adulthood increased as the frequency of bullying perpetration at 8 years of age increased. The associations between more frequent male bullying and the severity of violent offenses were assessed using multinomial logistic regression analyses. The explanatory variable was bullying perpetration and this had three categories: bullying never, sometimes or frequently. Violent offending was the outcome variable and this was also divided into three categories. The reference category was males who had never committed any violent offenses and the other two categories were those who had only committed minor violent offenses or those who had committed severe violent offenses. If an individual had committed a number of offenses, we recorded the most serious offense for this analysis. ORs and 95% CIs were estimated.
Second, the additional analyses also assessed whether the HRs would be different for just bullies and bully-victims. The explanatory variable had three categories and these were no involvement in bullying as a perpetrator at 8 years of age, just being a bully at that age or being a bully-victim. The HRs and 95% CIs were estimated using Cox regression analyses. The first event was the first violent offense of any severity.
Two-sided p values of less than 0.05 were considered statistically significant, except for the interactions, which were based on 0.1. The statistical analyses were conducted using SAS 9.4 for Windows (SAS Institute Inc, Cary, NC, USA).