Teun Remmers

Afterschool PA and the built environment using GPS, GIS and accelerometers | 133 Statistical analyses Days were used as the unit of analysis because this allows examining day-to-day variation within children. After presenting participant characteristics (Table 1), we present the median and interquartile ranges (IQR) for children's daily contribution to several domains and subdomains (Table 2) (51). In our multivariate explanatory analyses, we focused on the leisure time and the active transport domain. In the leisure time domain, we associations were separately analyses by LPA and MVPA. In contrast, in the active transport domain (i.e. cycling and walking), we combined LPA and MVPA intensity-categories. The first set of independent variables were meteorological variables, accessed from a local municipality weather station’s hourly registry. The second set of independent variables were variables from the multi-place environments, in which we focused on publically accessible environmental features accessed from the municipality's GIS registry. All multivariate analyses were performed using multilevel linear mixed models with a repeated term for days within children and a random term for school, accounting for the nested structure of measurement days within children and children within schools. Normality of residuals was inspected using normal probability plots. As all of our model-residuals showed significant deviation from normality, we transformed our dependent variables using log-transformations, while replacing all zero values with a minimal value of 0.01 square meters. Subsequently, model fit of these models were inspected and tested against the non-transformed variant to verify its fitting capabilities. To facilitate comparisons between independent variables, we calculated standardized log-transformed coefficients by dividing each log-transformed model-coefficient by its standard error. In analyses containing second level GIS-features of the built-environment, a manually executed stepwise procedure was followed to control and investigate the impact of potential multicollinearity issues. Namely, first we investigated associations between our dependent variables and all first-level GIS variables, adjusting for baseline variables (i.e. wear time, age, gender, and meteorology; see Table 3). Second, we replaced the first-level GIS variables with the accompanied second-level GIS variables, and tested these associations apart from each other. Third, we simultaneously entered second-level variables with strong associations ( p < 0.10), and deleted variables with the largest p- value; only retaining variables that were statistically significant ( p < 0.05). Baseline variables were not deleted from these models. An exception was made for the playground-variable (as this may be especially relevant for urban planning), from which we presented its association with leisure time PA regardless of its P-value. Moderation analyses were performed for gender by interaction terms and inspecting stratum-specific results, but significant interactions were not found. Likewise, we also investigated moderation for the daily minutes that respondents spent in their multi-place environment (based on their GPS location), but no such moderation was found. Statistical analyses were performed using SPSS 21.0 for Windows (IBM SPSS Inc., Armonk, NY), and p < 0.05 indicated statistical significance.