Crystal Smit

Chapter 4 76 For the main analyses, three structural path models were tested using Mplus 7.2 (Muthén & Muthén, 2012). The first model tested mean-level differences between conditions on water and SSB consumption after the intervention (T2 and T3), adjusting for consumption prior to the intervention (T1; see Figure 4.2a); the second model examined whether descriptive norms moderated the mean-level differences between conditions on subsequent water consumption and the third model examined whether injunctive norms moderated the mean-level differences between conditions on subsequent water consumption (see Figure 4.2b). In all models, condition was specified as two binary dummy variables with social network intervention as reference category (coded as 0). In the last two models, the social norm variables were centered prior to creating the interaction terms involving social norms and differences between conditions. The parameters in the models were estimated applying (full-information) maximum-likelihood estimation with robust standard errors (MLR in Mplus) to account for missing values and potential deviations from multivariate normality. Additionally, the models were adjusted for clustering of the sample—children were ‘nested’ in classrooms—using the Mplus procedure TYPE = COMPLEX, with classroom as the cluster variable. This procedure results in standard errors that are adjusted to account for non-independence within classrooms. The fit of the path models was assessed with the following good fit indices: root mean square error of approximation (RMSEA, with a cut-off value of < .08 and p -close > .05), comparative fit index (CFI, with a cut-off value of > .90) and normed c 2 ( c 2 /df, with a cut-off value of < 3.0; Hu & Bentler, 1999; Kline, 2011). In the structural path analyses, the unstandardized regression coefficient (b) provides the estimated mean-level difference between conditions on consumption behaviors following the intervention, adjusted for baseline consumption behaviors. Formodels yielding significant interaction effects, simple slope analyses (Aiken, West, & Reno, 1991) were used to examine the regression coefficient of the condition–consumption behavior relationship across two levels of the moderator (low social norms: -1 SD ; high social norms: +1 SD ).