We expect that many of the available treatments to treat AD have not been compared in any trial. In the absence of direct evidence for a comparison, indirect evidence can be obtained by combining trials that compare the interventions with a common comparator. By incorporating direct and indirect evidence in a single analysis, NMA provides an internally coherent set of relative treatment effects for all possible comparisons and, therefore, allows for a formal hierarchy of the interventions from the best to worst for each outcome [
30]. If transitivity assumption (i.e., similarity of included trials in terms of clinical and methodological characteristics that comprise important effect-modifiers) is deemed plausible, NMA can be safely applied to provide credible results [
31]. Otherwise, violation of transitivity assumption may cause incoherence between direct and one or more indirect effects beyond between-trial variance, and thus, reduce our certainty to the NMA results [
31]. We will assess the plausibility of transitivity assumption by investigating the distribution of important effect modifiers in each observed comparison [
32]. We will apply Bayesian RE NMA with consistency equation and incorporation of multi-arms trials to accommodate the anticipated statistically heterogeneity and to account for the correlation between treatment effects that share the same control arm in multi-arm trials [
27,
33]. Between-trial variance will be assumed common in the whole network to enable estimation of the parameter for comparisons with few trials, as information is “borrowed” by comparisons with many trials [
31]. Under this assumption, the correlation between treatment effects in multi-arm trials equals 0.5 [
27]. Coherence, the statistical manifestation of transitivity, will be investigated locally and globally. For the former, we will apply the node-splitting approach [
34,
35] using the R-package gemtc [
36] to automatically identify the comparisons to split in closed loops of interventions, whereas for the latter, we will compare the model-fit and complexity of the NMA model with and without consistency equation using the deviance information criterion (DIC) which provides a measure of model fit penalized for model complexity [
37]. The model with lower DIC will be considered to have a better compromise between model fit and complexity [
34,
38].
To illustrate the network geometry, we will create network plots for each outcome. The plots will display visual information of the evidence retrieved for each outcome regarding the number of trials and patients involved in each direct comparison.
In line with the statistical analysis for pairwise meta-analysis, we will use OR and mean difference or SMD as effect measures for binary and continuous outcomes respectively, and we will report the posterior mean for the treatment effects but the posterior median for
τ2 alongside their 95% credible intervals. Furthermore, we will extent the pattern-mixture model of Turner et al. [
28] for binary missing participant outcome data to operate in a network of interventions [
39], and we use the pattern-mixture model of Mavridis et al. [
29] to incorporate continuous missing participant outcome data as observed in the analyzed networks. We will create league tables for each outcome to present the NMA results for all possible comparisons as well as the results from pairwise meta-analysis for the observed comparisons with at least two trials. Furthermore, for each outcome, we will create forest-plots to present the NMA treatment effects alongside the respective direct and indirect treatment effects of comparisons with the reference intervention of the network. We will also present several measures of intervention hierarchy including the rank probabilities, ranks, and surface under the cumulative ranking curve (SUCRA) values [
40]. Specifically, we will create rankograms for each intervention and outcome to fully illustrate the uncertainty across the ranks [
40]. We will also present SUCRA plots to illustrate the cumulative ranking probabilities for each intervention and outcome [
40]; for each intervention, SUCRA value indicates the percentage of effectiveness (or safety) of that intervention as compared to an imaginary intervention that is always the best with certainty [
41]. For each outcome, the best treatment will have high SUCRA value and the worst treatment low SUCRA value. To aid the interpretation of the results in terms of hierarchy and relative treatment effects, we will incorporate the posterior median ranks and posterior mean SUCRA values alongside their 95% credible interval in the aforementioned forest-plots. We will create a scatter diagram to identify the best balance between efficacy and safety.