Addi van Bergen

Chapter 2 28 them as a demographic group rather than a high-risk group. Data extraction was performed by one reviewer (AvB, MB or KS) and checked by a second (AvB, BC or HS). Risk of bias assessment As there is currently little consensus on the critical elements for assessing risk of bias in observational studies [27], we opted for a two-track approach. The general methodological quality of each study was evaluated independently by two reviewers (AvB and MB KS BC or HS) using the Critical Appraisal Skills Programme (CASP) tools for cross-sectional and cohort studies (Supplementary files 2-3). The respective CASP checklists consist of 10 and 11 questions (e.g. “Was the outcome accurately measured to minimise bias?” and “Was the cohort recruited in an acceptable way?”), that can be answered with: ‘yes’ (1 point), ‘can’t tell’, or ‘no’ (0 points). The option to answer ‘yes moderately’ (0.5 points) was added by the reviewers. Disagreements were resolved through consensus and, if necessary, a third reviewer was consulted (BC). A commonly used cut-off point of 60% was used to distinguish between low and acceptable quality studies [28]. Only acceptable quality studies were included in the synthesis. As done by De Silva et al. [29], we assessed, in addition to the CASP, a number of specific methodological limitations with a high risk of bias for our research question. We examined whether the definition, operationalisation and measurement of SE/SI were adequately substantiated, whether testing of the association between SE/SI and health was a stated objective of the study and whether adjustment for confounding factors was performed. Details can be found in Supplementary file 4. Data analysis Given the variation in health measures and study designs, it was not possible to conduct a meta-analysis. Instead, we used the method of grouping results as originally described by Ramirez et al. [30]. To examine the six research hypotheses, we grouped the results for each hypothesis into four qualitative patterns. These were: 1) positive, when a significant (p < 0.05) concordant relationship was found for all measured SE/SI dimensions (high-SE/low SI corresponds to low health outcome), 2) negative, when an inverse association was found, 3) no association, when the relationships between the SE/SI dimensions and health were not statistically significant, and 4) partly (+/0), when studies reported multiple associations. We classified the result as partly when 30- 70% of the tested relations were positive and the remaining 70-30% not significant. If studies reported findings for multiple, non-overlapping, research groups, e.g. men and women, these were included separately in the data analysis and counted as separate instances. When both unadjusted and adjusted results were presented, only adjusted results were reported. Results were combined by counting the number of instances in each category and weighting by sample size.