We did all analyses separately for studies in children and adolescents and for studies in adults. First, we did pairwise meta-analyses (active drug vs placebo, or active drug vs another active drug) for all outcomes and comparisons at every available timepoint, using a random-effects model.35 We calculated standardised mean differences (SMDs), Hedges's adjusted g, and odds ratios (ORs), with relative 95% CIs, for continuous and dichotomous outcomes. We assessed statistical heterogeneity within each pairwise comparison by calculating the I2 statistic and its 95% CI.36 Second, we did network meta-analyses within a frequentist framework assuming equal heterogeneity parameter τ across all comparisons and accounting for correlations induced by multiarm studies.37, 38 We based the assessment of statistical heterogeneity in the entire network on the magnitude of the common τ2 estimated from the network meta-analysis models.39 We compared the magnitude of the heterogeneity variance with the empirical distribution.40, 41 We used the loop-specific approach42 and the design-by-treatment model43 to evaluate incoherence locally and globally, respectively. We established a hierarchy of competing interventions using surface under the cumulative ranking curve (SUCRA) and mean ranks.44
