Teun Remmers

72 | Chapter 5 Statistical analyses We first evaluated the association among outside play and all attributes of the perceived physical environment, and the social environment (parenting influences and social capital). To do this, we performed repeated measures linear mixed model analyses with outside play at age five and seven years as the dependent variable. We entered season of outside play measurement at age five and seven as a random factor. By doing so, we allowed each child to have its own random slope for season-combinations at five and seven years, while using an autoregressive (AR1) covariance structure. We examined all analyses using the following sequence: model 1) factor only adjusted for covariates (i.e. gender, maternal education, and child age); model 2) factor adjusted for covariates and all variables of their block (i.e. perceived physical environment versus parenting influences and social capital); 3) factor adjusted for covariates and all above described variables of the perceived physical environment, parenting influences and social capital (final model). We tested for moderation by entering interaction terms between each of the perceived physical environment variables, and each of the parenting influences and social capital variables, using the same repeated measures linear mixed model analyses as described above, with interaction terms for the moderators. We examined all potential moderating associations using the following sequence: model 1) interaction term only adjusted for main effects of the interaction and covariates; model 2) as 1, but also adjusted for previously defined statistically significant main effects; model 3) as 2, but also adjusted for previously defined non-significant main effects; 4) as step 3, but also adjusted for other statistically significant interaction terms in the model (final model). Finally, we stratified consistent significant interactions for interpretation purposes, using a median split. Based on results from previous studies (11, 25, 44, 45), we investigated the potential confounding influence of seasonality (autumn, winter, spring, summer), gender, age of the child, and maternal education. For maternal education, categories were low (no education, primary school, or ≤ 3 years of general secondary school), mid-low (< 3 years of general secondary school), mid-high (higher vocational training, undergraduate programs, or bachelor’s degree), and high (higher academic education) (46). None of these potential confounders were associated with a change of more than 10% in any of our coefficients after adjustment. However, to improve the precision of our models, these four variables were entered in our models as covariates (47). To compare the relative strength of associations among variables, we used standardized coefficients in all models. We defined statistically significant moderation as p < 0.10 and a statistically significant association for main effects as p < 0.05. As we performed various model variations to investigate consistent significant interactions, correction for multiple testing was not applied. All analyses were performed with SPSS version 20.0 (IBM Corp., NY, USA).