Teun Remmers

22 | Chapter 2 time period (arrow 1 in Figure 1), the child’s BMI z-score at T 0 was subtracted from the BMI z-score at T 1 , and divided by the time in years between the two BMI measurements. By doing so, the dependent variable reflected the mean change in BMI z-score per year, between T 0 and T 1 . The second time period (arrow 2 in Figure 1) was treated in the same way, but here the BMI z-score at T 1 was subtracted from the BMI z-score at T 2 . The fraction of time spent in each of the PA intensity categories was modeled separately. In the first period, relative PA intensity at T 0 was used and in the second period relative PA intensity at T 1 was used (see Figure 1). A proportion of the children ( n = 105) contributed to both time periods, depending on the availability of the measurements at each time point. To account for the correlation between these repeated time periods, we performed generalized estimation equation (GEE) analyses with an exchangeable correlation structure. The time period was used as the within-subject variable. In addition, GEE analyses assume that the effect of PA on BMI development is similar for the first and second time period. To test this assumption, the time period was used to test for statistical interaction with PA (time period x PA interaction). The most complete models were adjusted for origin of BMI measurement (research assistant/parents), cycling, swimming, season, paternal BMI, maternal BMI, and recruitment group (conventional/alternative). Because of optimal readability of Table 2 and realistic increments in PA in children (i.e. 5% roughly corresponds to 30 minutes per day), a PA increment of 5% was chosen to correspond to BMI z-score development. Randomness of missing values in the covariates (i.e. cycling, swimming, maternal BMI and paternal BMI) was checked by Little’s missing completely at random tests, and accompanying missing value analyses. Since none of the covariates with missing values showed strong deviations from the completely at random scenario, all missing values were imputed using multiple regression imputations. In total, 76 missing values of covariates were imputed. The percentage of imputed values per covariate at T 1 ranged between 0.6% for maternal BMI, and 10.7% for cycling and swimming. Analyses were stratified for weight status at the first measurement of the time period (i.e. initial weight status), as the effect of PA on the development of BMI was expected to be more favorable in children with a high initial weight status, compared to those with a low initial weight status. Due to statistical power arguments, relative lean weight status was defined as the 25 th percentile of our study sample and relative heavier weight status was defined as the 75 th percentile. Subsequently, children equal to or below the 25 th percentile were defined as ‘leaner’, and children equal to or above the 75 th percentile were defined as ‘heavier’. The effect of potential confounders was checked by inspecting the difference in unstandardized regression coefficients of the PA variable, in absence versus presence of the individual confounder. A confounding effect was considered when the unstandardized beta values differed by ≥ 10%, or when the confounder was considered to be essential to the study design (i.e. recruitment group). Potential effect modifiers (i.e. age x PA, gender x