Prognostic factors of adherence 61 Statistical heterogeneity If pooling was possible, we assessed the outcome on statistical heterogeneity by eye-balling. Second, we calculated the Cochran Q as the weighted sum of squared differences between individual study outcomes and the pooled outcome across all studies. When p was significant (<0.05), statistical heterogeneity was considered to be present [23]. Third, we used the I2 statistic to assess the variability between studies, the statistical heterogeneity. Low heterogeneity was considered with an I2 of less than 40%, moderate heterogeneity at 30%-60%, substantial heterogeneity at 50-90% and considerable heterogeneity at 75-100% [24]. Because we considered a random-effects model effect sizes could show more variance than when drawn from a single homogeneous population. This between-study heterogeneity was quantified by using t2. Subgroup- and sensitivity analysis If statistical heterogeneity was considered to limit the interpretability of the pooled effect estimate, we explored the heterogeneity using subgroup analysis. If a study was identified as an outlier after eye-balling of the forest plots, we investigated that study further. Based on this, other subgroup analyses could be considered a posteriori to explain the observed heterogeneity. Because we assumed a random-effects model within the subgroups, fixed-effects (plural) model (mixed-effects model) was considered. If the number of studies in the subgroups were small, we used a pooled version of t2 across all subgroups. To determine whether a statistically significant subgroup difference was detected, we considered a p-value for this test of less than 0.1 to indicate a statistically significant subgroup effect. Furthermore, to interpret the subgroup analyses, we used the criteria of Richardson et al [25]. We expected methodological heterogeneity in study design and methodological quality. Therefore, a sensitivity analysis was considered based on study design (‘observational studies’ vs. ‘other studies’) and/or risk of bias. We also considered a sensitivity analysis to examine what would happen if some aspect of the data or analysis were changed. [23]. Publication bias We used funnel plots and Egger’s test of small study effects [22] to assess the impact of possible publication bias when there were at least ten studies included in the meta-analysis. When there are fewer studies, the power of the test is considered too low to distinguish chance from real asymmetry. Funnel plots were assessed visually by eye-balling for asymmetry and statistically. A p-value <0.05 for the Egger’s test indicated substantial asymmetry of the funnel plots, thereby implying possible publication bias [22].
RkJQdWJsaXNoZXIy MjY0ODMw