2.6.2. Statistical Assessment

The network forest plot presented with MD and 95% CI of each intervention was used to rank each treatment strategy for visual and statistical verification. The P-score was also used to rank treatment, which assesses certainty that a specific intervention is better than competing inventions. The P-score is nearly identical to the numerical values of SUCRA in the Bayesian model NMA [33]. For the consistency assumption, we checked both global (network level) and local approaches (particular contrast of intervention level) [21]. In the

global approach, we used the 'decomp.design' function of R software to assess consistency under the assumption of a full design-by-treatment interaction random effect model [34]. Q statistics were used to assess inconsistency in the global approach. If the *p*-value for the Q statistics was below 0.05, it was assumed that significant inconsistency (disagreement) existed in the global network. In the local approach, we adopted the net-splitting method to split the network estimation of the effect size on each intervention into direct and indirect evidence using the Facenetsplit function of R software. It calculates the difference between direct and indirect estimates and assesses whether the difference is statistically significant [34]. Net-split plots were also provided for visual inspection of inconsistencies between direct and indirect comparisons. If the *p*-value for the net-split analysis was below 0.05, it was assumed that significant inconsistency (disagreement) existed in a specific local loop, which indicates a considerable difference between indirect and direct effect size estimation. If there were significant disagreements in the local or global approach, we conducted a sensitivity analysis by sequentially excluding studies one by one. If we identified which studies were inconsistent, we excluded studies from the NMA. A net league table is also presented. The upper right triangle presents the effect size estimated by only direct comparison, which is similar to the pairwise comparison. As direct comparison does not exist in all treatment comparisons, there are several blanks in the upper triangle. The lower left triangle provides a pooled estimation of the direct and indirect comparisons of the effect size.
