Esther Mertens

54 | Chapter 3 was poor (Cronbach’s α = .53 - .60). The 3 items that included a negative (“I don’t think a lot about my thoughts.”, “I almost never participate in ‘self-reflection’.”, and “I don’t think about the reason why I behave the way I do.”) were deleted to avoid a double negative and to improve reliability (Cronbach’s α = .74 - .89). Self-esteem. Students’ level of global self-worth was measured using the subscale Global self-perception of the Self-Perception Profile (Harter, 1988) completed by the students. The subscale has 5 items (e.g., “I am satisfied with myself.”) answered on a 4-point Likert-type scale (1 =  completely not true to 4 = completely true ). Some items were recoded so that higher scores indicated high levels of self-esteem (Cronbach’s α = .73 - .75). Emotional self-regulation. Students completed the Difficulties in Emotion Regulation Scale (Anderson, Reilly, Gorrell, Schaumberg, & Anderson, 2016) to assess students’ abilities to control their emotions and their access to emotion regulation strategies (e.g., “When I’m upset, I know that I can find a way to eventually feel better.”). The questionnaire consists of 14 items answered on a 5-point Likert-type scale (1 = almost never to 5 = almost always ). Some items were recoded so that high scores indicate higher levels of emotional self-regulation (Cronbach’s α = .88 - .91). Statistical Analyses Data were analyzed using an intention-to-treat approach in which students assigned to the intervention were included in the analyses regardless of whether they actually participated in the intervention or not. Participants were nested in schools in classes. We took clustering at school level into account by applying the complex sample cluster feature of M plus (version 8.2; Muthén &Muthén, 2010). This is a conservative clustering procedure providing unbiased estimates of the standard errors (Muthén & Muthén, 2010). Clustering at class level was not taken into account as class composition was not stable over the years (e.g., Cross et al., 2016). To include all participants in the model, we used Full Information Maximum Likelihood (FIML) procedures. Parameter estimates were obtained through Robust Maximum Likelihood estimation (MLR) which is robust to non-normality and non-independence of observations (Muthén & Muthén, 2010). To examine the effectiveness of R&W, we tested a series of latent growth curve (LGC) models in M plus , as suggested by Greenberg and Abenavoli (2017). LGC models estimate for each participant an individual growth curve based on his/her initial level (i.e., intercept) and change over time (i.e., slope). The individual growth curves are indicators of latent variables describing average group growth trajectories allowing for differences in trajectories between participants (Muthén & Muthén, 2010). The slope is of main importance; when the intervention is effective compared to the Control condition, it significantly alters the slope in the desired direction. To allow for nonlinear growth, we did not specify the rate of growth at T2 and T3 (Duncan & Duncan, 2004). Growth rates at T1 and T4 were specified at respectively 0 and 3.