Effects of the KiVa Antibullying Program on Cyberbullying and Cybervictimization Frequency Among Finnish Youth

Abstract

Cyberbullying among school-aged children has received increased attention in recent literature. However, no empirical evidence currently exists on whether existing school-based antibullying programs are effective in targeting the unique aspects of cyberbullying. To address this important gap, the present study investigates the unique effects of the KiVa Antibullying Program on the frequency of cyberbullying and cybervictimization among elementary and middle school youth. Using data from a group randomized controlled trial, multilevel ordinal regression analyses were used to examine differences in the frequencies of cyberbullying and cybervictimization between intervention (N = 9,914) and control students (N = 8,498). The effects of age and gender on frequencies of cyber behaviors were also assessed across conditions. Results revealed a significant intervention effect on the frequency of cybervictimization; KiVa students reported lower frequencies of cybervictimization at posttest than students in a control condition. The effect of condition on the perpetration of cyberbullying was moderated by age. When student age was below the sample mean, KiVa students reported lower frequencies of cyberbullying than students in the control condition. We also found evidence of classroom level variation in cyberbullying and cybervictimization, suggesting cyberbullying is in part a classroom-level phenomenon. KiVa appears to be an efficacious program to address cyber forms of bullying and victimization. We discuss several unique aspects of KiVa that may account for the significant intervention effects. Results suggest that KiVa is an intervention option for schools concerned with reducing cyberbullying behavior and its deleterious effects on children's adjustment.

Notes

Note. Intraclass correlation coefficients (ICCs) were tested for significance via deviance tests (Snijders & Bosker, 2012). ΔD = difference in model deviances (−2loglikelihood); p = probability value for deviance test.

Note. N = number of students endorsing variable; p = proportion of students endorsing variable; CV1 = cybervictimization at baseline; CV3 = cybervictimization at posttest; CB1 = cyberbullying at baseline; CB3 = cyberbullying at posttest; TRV = traditional victimization; TRB = traditional bullying; AGE = age of student; BOY = gender of student (female = 0, male = 1); INT = assignment to study condition (0 = control, 1 = intervention); SWE = language of instruction (0 = Finnish, 1 = Swedish).

¹Two “state-of-the-art” methods for dealing with missing data (Enders, 2010) were considered: full-information maximum likelihood (FIML) estimation and multiple imputation. We chose to use FIML which, in the multivariate case, uses all available information in the dataset to estimate model parameters. With a single-dependent variable, however, FIML will remove cases that have missing data on any variables—similar to the listwise deletion method—unless distributional assumptions (e.g., normality) are made for some of the predictors. Such assumptions were not tenable for our set of predictors. Although this is not optimal, there is not yet consensus on how best to handle multiple imputation of categorical data (Gebregziabher & DeSantis, 2010). Ultimately, we decided that the loss of power and precision would not be significant given the large sample size in the current study and that the small differences between responders and dropouts on observed baseline variables further supports our use of FIML.

Note. Odds ratios (OR) and associated confidence intervals for negative parameter estimates were converted to be above 1.0 for ease of comparison (Osbourne, 2006). In such circumstances, the direction of interpretation must be inverted. Est = parameter estimate; SE = standard error; CI = confidence interval; Res. = Residual.