Xuxi Zhang

concurrent validity have been widely applied by previous studies. However, there is no ‘golden standard’ of frailty to be used as alternative measure of the TFI. The alternative measures for psychological and social domains was limited by the available data in the UHCE project. In Chapter 3 , physical activity was measured by one question, instead of a validated multi ‐ item instrument such as the International Physical Activity Questionnaire (IPAQ) 41 . The item that we applied could not differentiate between type of activity and does not take the duration of activity into account. However, some previous studies 42 ‐ 44 indicate that using a single question to measure physical activity is appropriate in certain circumstances, e.g. when the sample size is large, when more complex methods would add to respondent burden, and when collecting data from a broad range of settings. Grill et al. (2012) also suggest that the reliability and validity of a single question to briefly classify physical activity levels is acceptable. 45 Secondly, we transferred the ordinal variable of physical activity into a dichotomous one; however, we conducted additional analyses on the association between physical activity and frailty with the ordinal variable of physical activity and found similar results. Lastly, there may be overlap between physical activity and two items of the TFI (walking and balance) which could cause over ‐ estimation of the association. However, when we explored the association between physical activity and overall frailty, after deleting these two items the results were similar. Statistical analysis Confounding and moderation Confounding variables are associated with both the determinant and the outcome under study, but are not on the causal pathway. 46 Ignoring confounding variables could lead to an overestimation or underestimation of the true association between the determinant and the outcome. 46 In all studies in this thesis, with the exception of Chapter 4, 6 and 7, we adjusted for potential confounders, which were carefully chosen based on previous literature, availability in the data and exploratory analyses. However, the possibility of residual confounding due to unmeasured or insufficiently measured determinants cannot be ruled out. ‘Moderation’ happens when the association between the determinant and the outcome varies according to a third variable. 47 In Chapter 2, 3 and 8 , we tested moderation by formal interaction tests and stratified data when there was significant interaction. We applied the Bonferroni corrections for interaction testing in case of multiple testing to avoid ‘chance findings’. Meta ‐ analysis (Chapter 6 and Chapter 7) “Any kind of variability among studies in a meta ‐ analysis may be termed heterogeneity”. 48 Clinical heterogeneity includes variability in the participants, interventions and outcomes, and methodological heterogeneity includes variability in study design and risk of bias. 48 Variability in the intervention effects being evaluated in the meta ‐ analysis is known as statistical heterogeneity, and is a consequence of clinical or methodological heterogeneity, or both, 228 Chapter 9