If appropriate, a Bayesian NMA will be conducted to synthesize results for each outcome. Nodes in the NMAs will represent the various geriatrician-led care models. CGAs are complex interventions that are often tailored to individual needs and implemented in different healthcare settings. We will define nodes to lessen the potential heterogeneity in the network. The role of the geriatrician and the healthcare setting of the CGA are known factors that contribute to the diversity of CGAs. Geriatricians will be consulted during node categorization to ensure that similar geriatrician roles are grouped together, which will aid in the clinical interpretation of results. Depending on the breadth of included studies and the extent of heterogeneity examined by simple classification, we may also consider dividing the nodes further into components and examine the members of the multidisciplinary teams and the frequency of contact with the multidisciplinary team and geriatrician.
For NMA results to be valid, assumptions of comparability, such as connectivity, transitivity, and consistency need to be considered [
36]. Connectivity refers to how well the studies form a connected network and will be visualized using network diagrams [
37]. Although there is no formal definition of what constitutes a connected network, the following will be used as a guide; a connected network should link the different interventions from each study and enable the connection of at least two interventions in the evidence base, and for every two studies, there should be at least one common comparator. Sparse or unconnected networks are not suitable for NMA, due to the heavy reliance on indirect evidence and will be assessed on their network geometry [
38]. Once connectivity is observed, transitivity will be explored, which allows us to make indirect comparisons by way of a common comparator (i.e., standard care) [
36]. To assess whether indirect comparisons are valid, we will color the edges in the NMA diagram to represent patient and study characteristics and assess visually whether there is an equal distribution of these characteristics across the interventions [
36]. Patient and study characteristics that we anticipate may influence transitivity include frailty, comorbid conditions, gender, and healthcare setting (i.e., acute care, long-term care settings). The consistency of the results from direct evidence (i.e., head-to-head trials from pairwise MA) versus indirect evidence (i.e., from NMA) will be compared using the design-by-treatment model [
39], which assesses global inconsistency across the network. If global inconsistency is observed, we will check for data abstraction errors and if none exist, we will explore local inconsistency using the loop-specific method [
40]. If inconsistency is still observed, we will conduct meta-regression and/or subgroup analysis. Potential meta-regression or subgroup analyses that we may explore will include age, frailty, cognition, functional status, attrition (<10 versus ≥10%), care setting (acute care, primary care, long-term care settings), and risk of bias (high versus low).
All NMAs will be conducted in OpenBUGS [
32] Bayesian statistical software. Results will be reported as odds ratios for dichotomous outcomes or mean difference for continuous outcomes, along with 95% credible intervals based on 100,000 Markov chain Monte Carlo simulations after a burn in of at least 50,000 simulations and vague priors. Model convergence will be assessed using the trace and history plot functions in OpenBUGS, as well as the Brooks-Gelman-Rubin (BGR) statistic [
34]. Predictive intervals will be calculated, which gives an indication of how the results will change if a new trial is conducted in the future and added to the evidence. The effectiveness of care models will be ranked using surface under the cumulative ranking (SUCRA) curve [
37]. Sensitivity analysis will be conducted to examine the effects of imputations for missing data [
31].