Teun Remmers

General discussion | 187 Ecological frameworks highlight interactive relationships between determinants and PA. This means that when interested in relationships between determinants and PA, third variables may play a role. First, a third variable may act as a confounder. Confounding variables bias the underlying relationship, as the variable is related to both the determinant and PA, while not being an intermediate factor in the causal pathway between the determinant and PA. Although it is important to adjust for potential confounders in multivariate models, residual confounding of unmeasured factors (or poorly measured variables due to misclassification) may have played a role in this thesis. Second, a third variable may act as a moderator or effect modifier. In this case, the relationships between the determinant and PA differs for the moderating variable. For example, in this thesis, we found some evidence that relationships between PA enjoyment and PA behavior differed by gender and impulsivity. In addition, we also found that children's responses to certain weather elements differed by type of the day (i.e. weekdays versus weekend days). Third, a third variable may act as a mediator. Mediating variables are related to both the determinant and PA, and are in between the causal pathway. Since in this thesis only observational studies were performed, we cannot speculate on potential causal pathways. Therefore, we did not speculate on potential mediating mechanisms in this thesis. Consequently, researchers are encouraged to also consider non-linear relationships and non-normally distributions in predicting PA patterns, and to investigate possibilities of somewhat more dynamic approaches that test multiple mediating or moderating pathways such as agent-based models (88, 89) or structural equation models (90, 91). The increased demand on analyses investigating complex pathways explaining specific PA contexts often require subgroup analyses or time-segmented data. Moreover, as devices for PA measurement (e.g. accelerometers) still produce random errors and selective data loss is still an issue, this in turn may lead to power issues. Future studies should account for performing these refined analyses (e.g. afterschool time-segments) and measurement error in their power calculation. Generalizability A number of studies presented in this thesis used data from the KOALA Birth Cohort study (chapters 2, 3 and 5). This cohort study started in 2000 with the recruitment of healthy pregnant women from the general Dutch population, with a focus on the development of children's asthma and allergies, and growth and development. The samples included in our analyses consisted of relatively highly educated mothers (91% at least mid-high education) (92, 93), with relatively low percentages of non-Western ethnicities (94). In addition, chapters 8 and 9 used data from the PHASE-kids study. This study started in 2015 and recruited children in their last year of primary school, to investigate development of PA in the transitional phase to secondary school. Again, parents of participating children were relatively highly educated (44% of responding parents had at least a higher vocational diploma). This is comparable with rates of the Dutch general population of that age (95). Of the 30 primary schools that were invited to participate, 21 primary schools were included in the study. Also, approximately half of the recruited