a. Reviewer Information

Both the Therapeutic Review of Clinical Trials and Review in Brief will be prepared by the DSEN Collaborative for Network Meta-Analysis in consultation with two external clinical experts specializing in cardiology.

b. Development of Research Questions

The research questions were developed jointly by jurisdictions, expert committee members, clinical experts, and clinical reviewers in consultation with pharmacoeconomic reviewers.

c. Literature Search Methods

The literature search will be performed by an information specialist. The following bibliographic databases will be searched through the Ovid interface: MEDLINE, MEDLINE In-Process & Other Non-Indexed Citations, Embase, DARE (Database of Abstracts of Reviews of Effects) for November 2011 and the Cochrane Central Register of Systematic Reviews, Controlled Trials and Economic Evaluations. Methodological filters will be applied to limit retrieval to RCTs and NRSs. The search will be restricted to clinical articles published from 1980 to December 2, 2011. Where possible, retrieval will be limited to the human population. Grey literature (literature that is not commercially published) searched to December 8, 2011 were identified by searching the websites of health technology assessment and related agencies, guideline producers, and professional associations that maintain safety information on pharmaceutical products.

d. Study Selection

Each review author will independently select studies for inclusion in the review according to the predetermined selection criteria. Review authors will independently make the final selection of studies to be included in the review, and differences will be resolved through discussion. A list of included and excluded studies will be listed in the report.

RCTs and NRSs will be selected for inclusion if they were published in English, assessed the desired intervention and comparators, reported relevant outcomes, and involved patients with non-valvular atrial fibrillation.

e. Quality Assessment

Quality assessment of RCTs will be conducted independently by two reviewers, with a third reviewer used to resolve disputes. Assessment of study quality for RCTs will be performed using the SIGN 50 instrument for RCTs for interval validity and the Cochrane Collaboration’s Tool for Assessing Risk of Bias. In addition, for NRS, the SIGN 50 instrument for cohort studies and an adaptation of the Cochrane Collaboration’s Tool for Assessing Risk of Bias will be applied.

f. Data Analysis

Two levels of analyses will be conducted. First, a meta-analysis will be conducted (RevMan 5.0 software from the Cochrane Collaboration will be used) and heterogeneity will be assessed leading to a fixed or random effects model or no analysis being considered. For each outcome, the weighted odds ratio and corresponding 95% confidence interval will be calculated for the overall treatment effect. At the next level, network meta-analyses (NMAs) will be conducted. Two approaches will be considered: the Bayesian mixed-treatment comparison model and the frequentist general linear mixed model. Following careful assessment of heterogeneity across trials in terms of patient characteristics, trial methodologies, and treatment protocols, NMAs will be conducted for the pre-specified outcomes. The effect estimate chosen (e.g., odds ratio, relative risk, or hazard ratio for stroke) will depend on the outcome of interest and the availability of data. For reference case NMAs, comparators considered will be dabigatran, rivaroxaban, apixaban, adjusted-dose warfarin, fixed low-dose warfarin, and fixed low-dose warfarin plus aspirin; however, some comparators (e.g., dabigatran) may be stratified by dose.

For Bayesian NMA, both fixed and random-effects models will be conducted; model selection will be based on the DIC and residual deviance. R (R Foundation for Statistical Computing, Vienna, Austria) and WinBUGS (MRC Biostatistics Unit, Cambridge, UK) will be used for Bayesian network meta-analyses according to the routine which accommodates evidence structures (which may consist of multi-arm trials as developed at the Universities of Bristol and Leicester (www.bris.ac.uk/cobm/research/mpes/). Adjusted-dose warfarin therapy will be the reference group for all Bayesian NMAs. Posterior densities for unknown parameters will be estimated using MCMC methods. Basic parameters will be assigned non-informative or vague prior distributions; more informative priors may be considered after evaluation of the information base and clinical expert advice. Point estimates and 95% credible intervals will be used to summarize findings. The probability of a comparator being optimal will be estimated for each outcome based on the proportion of MCMC simulations in which its relative measure of effect was best. The mean rank for each comparator will also be calculated. Consistency between direct and indirect evidence will be formally assessed using back-calculation and node-splitting techniques.⁶⁷ Graphical methods and numerical summaries will be developed for presenting results from network meta-analysis.⁷¹ Model diagnostics will also include trace plots and the Brooks-Gelman-Rubin statistic⁷² to assess and ensure model convergence. Two chains will be fit in WinBUGS for each analysis, each usually employing ≥ 20,000 iterations, with a burn-in of ≥ 20,000 iterations. Whether novel agent effects are present will be examined and their magnitude of effect estimated.^70,71

For frequentist NMA, a GLMM will be used. According to the outcome of interest, the GLMM model that follows the different distribution will be conducted (e.g., if the outcome follows a binomial distribution, then a mixed log-binomial model will be employed, with the log link function to generate the relative risk or with the logit link function to generate the odds ratio estimates). Adjusted-dose warfarin therapy will be the reference group for all NMAs. For reference case NMAs, comparators considered will be dabigatran, rivaroxaban, apixaban, adjusted-dose warfarin, fixed low-dose warfarin, and fixed low-dose warfarin plus aspirin; however, some comparators (e.g., dabigatran) may be stratified by dose. The random effect GLIMMIX model will be conducted. Two random effects can be considered in the model. The random effects trial accounts for the response variables of patients within a given trial being correlated. The random effects trial-by-treatment accounts for the correlation of responses between any two patients from the same treatment arm within a given study. The GLIMMIX procedure in SAS/STAT (SAS Institute Inc., Cary, North Carolina, USA) will be used for GLMM NMAs. Point estimates and 95% confidence intervals will be used to summarize findings. The rank for each comparator being the best treatment will also be calculated. Consistency between direct and indirect evidence will be formally assessed using back-calculation.⁶⁷ Graphical methods and numerical summaries will be developed for presenting results from network meta-analysis.^70,71 Model diagnostics will be evaluated using the diagnostic plots (e.g., residual plots) to assess and ensure model convergence.

For both the Bayesian and frequentist approach, and provided sufficient data are available to inform the evidence network, we will conduct meta-regression and/or subgroups analyses to adjust for CHADS₂ score, TTR, age, year of study, length of follow-up, gender, and history of a stroke to test the robustness of reference case analyses. In other sensitivity analyses, we will remove studies from the network that are of poor methodological quality.

Subgroups of interest include: time spent within the therapeutic range (TTR), CHADS₂ (or CHA₂DS₂-VASc) score, age (stratified as < 65 years, 65 years to 74 years, and ≥ 75 years), weight, impaired renal function (including mild, moderate, and severe), prior history of GI bleed, concurrent use of antiplatelet agents, concurrent use of NSAIDs.

If relevant heterogeneity is present, sensitivity analysis will be conducted based on aspects of the PICO statement and study methodology. Reporting bias will be assessed by constructing funnel plots for each outcome.

g. Writing of the Review Report

The review report will be written by the DSEN Collaborative for Network Meta-Analysis, with input from two clinical experts. A detailed internal review of the review report will be undertaken. Comments will be received from peer and external reviewers.

a. Clinical Systematic Review

Selection Criteria

A study was included if it met all of the inclusion criteria and none of the exclusion criteria summarized in Table 39.

Table 39

Inclusion and Exclusion Criteria for Primary Studies.