اثر متفاوت برنامه مقابله با زورگویی KIVA در زورگویان محبوب و منفور

Abstract This study utilized data from the evaluation of the Finnish KiVa program in testing the prediction that school bullies' high perceived popularity would impede the success of anti-bullying interventions. Multiple-group structural equation modeling (SEM) analyses were conducted on a subsample of 911 third-, fourth-, and fifth-graders identified as perpetrators of bullying. They belonged to 77 Finnish schools, including 39 schools implementing the KiVa program and 38 control schools. Data on peer-reported bullying and perceived popularity were collected before program implementation and one year later. Controlling for sex, age, and initial levels of bullying, KiVa participation resulted in lower rates of bullying (indicated by fewer peer nominations) after one year for bullies of low and medium popularity. However, there was no significant effect for those high in popularity, suggesting that popular bullies are less responsive to anti-bullying interventions than less popular bullies.

Results Analytical strategy We conducted multiple-group SEM analyses on the subsample of 911 bullies. A robust maximum-likelihood estimation method was used to account for the nested data structure (i.e., bullies nested in classrooms). Bullying at both time points was modeled as a latent factor, with correlated residuals estimated for corresponding indicators at T1 and T2. Prior to evaluating the effects of KiVa participation, we tested for measurement invariance across time (T1 and T2) and group (low, medium, high popularity) using well-known procedures and criteria (Brown, 2006 and Meredith, 1993). A sequence of models was fit, from an unconstrained model (i.e., configural invariance model) to a more restricted model, imposing equality constraints on the corresponding factor loadings (i.e., weak invariance model). The resulting change in model fit was evaluated by comparing CFI and RMSEA indices (Cheung and Rensvold, 2002 and Little et al., 2007). In the primary analysis, bullying at T2 was predicted by KiVa intervention status, controlling for sex, age and bullying at T1 (see Fig. 1). Bullying at Time 2 (T2) as predicted by KiVa intervention status, age, sex and ... Fig. 1. Bullying at Time 2 (T2) as predicted by KiVa intervention status, age, sex and bullying at Time 1 (T1) within low, medium, and high popularity groups. Residual variances and correlated residuals were included in the model, but are not depicted for clarity. Figure options Some of the variables included in the analyses had missing data. The proportion of missing data was 11.7% for age and 2.3% for popularity. For bullying at T1, only the third item (“always finds new ways of harassing the victim”) had some missing data (10.8%); these missing values are considered to be missing at random, as the missingness was due to a technical problem in the administration of the survey resulting in 36 classrooms (in the full sample, 33 classrooms in the subsample of bullies) not being presented with the item. For bullying at T2, the three items had 21.8% of missing data, which were primarily due to entire classrooms dropping out of the study. A t-test indicates that both T1 bullying and popularity scores do not differ among bullies missing at T2 versus bullies not missing at T2 (bullying: t = .240, p = .810; popularity: t = .021, p = .983). The data can also be considered missing at random. Analyses were conducted using Mplus v.6.1 (Muthén & Muthén, 1998-2010), which utilizes full information maximum-likelihood procedures to account for missing data considered missing at random (MAR). These procedures use all available data in generating the final parameter estimates, which are not biased by the proportion of missing data in our study (Enders, 2010 and Graham et al., 2007). Measurement model We tested measurement invariance for the bullying factors across time (T1 and T2) and group (low, medium, high popularity). The initial, freely estimated model had excellent fit, χ2(15) = 23.20, RMSEA = .043 (95% CI = .000, .075), CFI = .995, TLI = .985, SRMR = .022. Following standard procedures to evaluate measurement invariance, we equated factor loadings across both time and group and found no significant changes in fit based on the RMSEA (i.e., the nested model RMSEA was included within the 90% RMSEA confidence interval of the constrained model; see Little et al., 2007) and CFI (i.e., changes in CFI were less than .01, see Cheung & Rensvold, 2002). This result shows that the bullying constructs, as measured by the three items, were qualitatively equivalent across school years and popularity groups (i.e., the same bullying constructs are being assessed in students of low, medium, and high popularity, and at T1 and T2). The standardized factor loadings for the bullying measures ranged from .51 to .83 (ps < .001) at T1 and from .87 to .93 at T2 (ps < .001) across low-, medium- and high-popularity groups. Residual variances were generally small to moderate (from .15 to .47, with a maximum of .74); residual covariances between corresponding indicators were low (from .01 to .28). Overall, the measurement model estimates show that the bullying indicators are valid representations of the underlying latent constructs. Structural model The model had good fit, χ2(77) = 180.04, RMSEA = .067 (95% CI = .054, .080), CFI = .948, TLI = .926, SRMR = .110, based on existing guidelines (Hu & Bentler, 1999). Results are presented in Table 1. KiVa participation resulted in lower proportions of peer nominations for bullying behavior for bullies in the low-popularity group (p = .035) and in the medium-popularity group (p < . 001). The effect size for the KiVa intervention effects was small to moderate, as reflected by the standardized estimates (see Cohen, 1988). However, KiVa participation did not significantly decrease the proportion of peer nominations for bullying behavior for bullies in the high-popularity group (p = .740). Comparing the KiVa intervention effect across popularity groups resulted in the following: the effect of KiVa participation on bullying was significantly stronger in the medium popularity group compared to the high popularity group, Satorra–Bentler 2 scaled χ2(1) = 6.05, p = .014; however, the strength of the KiVa effect in the low popularity group did not differ from that in the medium popularity group, Satorra–Bentler scaled χ2(1) = 1.59, p = .207, or high popularity group, Satorra–Bentler scaled χ2(1) = 2.06, p = .151. Table 1. Effects of KiVa intervention on bullying at T2, controlling for age, sex and bullying at T1 for low, medium and high popularity groups (N = 911). Unstandardized estimate S.E. Standardized estimate Low popularity (N = 267) KiVa intervention − 0.182* 0.086 − 0.146 Age − 0.068 0.076 − 0.054 Sex (boy) 0.092 0.072 0.073 Bullying T1 0.708*** 0.148 0.567 Medium popularity (N = 290) KiVa intervention − 0.330*** 0.093 − 0.231 Age − 0.013 0.100 − 0.009 Sex (boy) − 0.094 0.142 − 0.066 Bullying T1 0.989*** 0.110 0.693 High popularity (N = 335) KiVa intervention − 0.030 0.089 − 0.023 Age − 0.246** 0.096 − 0.188 Sex (boy) − 0.012 0.091 − 0.009 Bullying T1 0.876*** 0.132 0.668 Note. Popularity data were missing for 19 of the 911 bullies selected for the analyses. *p < .05. **p < .01. ***p < .001. Table options Across popularity groups, bullying scores at T1 strongly and positively predicted bullying scores at T2 (ps < .001); the higher the proportion of peer nominations received for bullying behavior before the start of the program, the higher the proportion of peer nominations received for bullying behavior after one year of program implementation. Age and sex were not significant predictors of bullying at T2 in the low-popularity group (p = .372 and p = .201, respectively), nor in the medium-popularity group (p = .897 and p = .507, respectively). In the high-popularity group, age had a significant effect on T2 bullying (p = .010); older children received lower bullying scores. There was no effect of sex (p = .894).