154 6 Chapter 6 7. Other bias. As cluster-randomised trials were also included, we added the following design-related domains for these types of studies. 1. Recruitment bias (i.e. whether individuals were recruited after clusters had been randomised). 2. Baseline imbalance between groups (i.e. whether risk of baseline di erences was reduced by using strati ed or pair-matched randomisation of clusters). 3. Loss of follow-up of clusters (i.e. whether missing clusters and missing outcomes for individuals within clusters could lead to a risk of bias in clusterrandomised trials). 4. Methods of analysis adequate for cluster-randomised controlled trials (i.e. whether clustering was taken into account in the analysis) (Higgins 2011). 5. We judged all items as having high, low, or unclear risk of bias and provided a quote from the study and/or a justi cation for our decision. Measures of treatment e ect We analysed results of the studies in RevMan 5, using random-e ects modelling. We used forest plots to compare results across trials. When possible, results were related to the minimum clinically important di erence (MCID) for the respective variable. We undertook meta-analysis only when this was meaningful, that is, when treatment, participants, and the underlying clinical question were similar enough for pooling to make sense, and when the results of at least two RCTs were available. We used intention-to-treat data or the ‘full analysis set’ whenever reported. We used per-protocol analysis when neither was reported. Normally, outcome measures that have been adjusted for baseline di erences produce the most reliable outcomes. However, these can be analysed only by generic inverse variance (GIV). Also, we noted signi cant variation in the number of parameters adjusted for between studies. Hence, we used unadjusted values in our random-e ects modelling for studies with an RCT design, and values adjusted for potential clustering e ects for studies with a clusterRCT design. When multiple trial arms were reported in a single study (e.g. hospital-based pulmonary rehabilitation and home-based pulmonary rehabilitation), we included all relevant trial arms. We halved the control group in these cases to avoid double-counting, as suggested in the Cochrane Handbook for Systematic Reviews of Interventions (Chapter 16.5.4) (Higgins 2019a). Unit of analysis issues When a study used a cluster-RCT design, we calculated the estimate of e ect by using the GIV whenever possible. We used the mean di erence (MD) and the 95% con dence interval (CI) reported by study authors when the appropriate analyses were used and authors had adjusted for cluster e ect. We calculated a dummy mean change and standard deviation (SD) based on the MD and its 95% CI for cluster-RCT studies, as
RkJQdWJsaXNoZXIy MTk4NDMw