Teun Remmers

186 | Chapter 10 Selection Bias and Information Bias Complete data-collection and prevention of selective loss to follow-up is of particular interest in PA studies, as accelerometer protocols require multiple days of measurement. In addition, most accelerometer data-reduction protocols forces researchers to exclude high amounts of data because of insufficient wear time. Missing data may be at random, or associated with participant characteristics. When missing values occur in a random pattern (i.e. unrelated to any outcome or determinant), they are expected to only reduce the power and precision of the associations. In this case, multiple imputation (MI) procedures are valid in preserving the power and precision of associations of interest. On the other hand, when missing values occur in a non-random pattern, relationships may be significantly biased, and MI-procedures are not recommended. In this thesis, we observed that children's compliance to wear accelerometers was reduced with increased wear time. This was especially the case with the heavier and slightly more bothersome GPS loggers. Moreover, consistent with observations by Bürgi et al. (2015) and van Kann et al. (2016), participants sometimes mentioned that they deliberately removed the equipment while at a sports club, as they perceived the devices as uncomfortable (85, 86). This may potentially lead to selective loss of data, and therefore may bias associations at hand. Researchers may account for potential random loss of data due to decreased children's compliance by oversampling studies at baseline (61). Also, selection bias may occur in the recruitment of participants in these studies, as GPS loggers collect private or sensitive information. Therefore, these studies are encouraged to communicate transparent protocols regarding data analysis, reporting of results, anonymity, and storage of data. Technological advances in PA measurement are developing quickly. For example, with recent innovations in wrist-worn accelerometers (43), or smartphone applications (87) children may be more willing to wear equipment for multiple consecutive days, and to keep wearing the devices at sports clubs. However, prior to applying potentially less bothersome devices in population-based studies, assessing practical issues and validity of these fast-developing innovative devices is necessary (43). Data Analyses In this thesis, predominantly linear models were used, often accounting for multilevel clustering of seasonality/weather elements (chapters 3, 4, and 5) and children within schools/classrooms (chapters 8 and 9). However, as seen in the relationship between temperature and PA (chapter 6), relationships may not always be linear. Therefore, researchers are encouraged to investigate linearity and to apply non-linear models where appropriate. In addition, as described in chapters 8 and 9, minutes that children spent in specific contexts were often not normally distributed. Although log-transformations improve model-fit and help to meet the model assumptions, comparisons between independent variables and straightforward interpretation of the effect size is often difficult. Future researchers are encouraged to explore potential non-parametric testing in non-normally distributed PA data, in order to meet model assumptions and to preserve power and precision.