Article type
Year
Abstract
Background: If several interventions are available for the same condition, randomized trials provide a network of evidence comparing different pairs of interventions, which is difficult to interpret.
Objectives: To present a combination of frequentist and Bayesian approaches towards exploring the consistency of the network and summarising direct and indirect comparisons in a single coherent and understandable analysis.
Methods: Using an example from interventional cardiology, we first conducted a standard random-effects meta-analysis separately for randomized comparisons of three pairs of interventions. Then, we used random-effects meta-regression to perform indirect comparisons of different interventions from trials that had one intervention in common. As a measure of inconsistency we calculated the I2 statistic. To combine direct within-trial comparisons with indirect between-trial comparisons, we used a frequentist approach to network meta-analysis, assuming two levels of random variation, heterogeneity and sampling error at the level of individual trials and incoherence at the level of pairs of interventions. The incoherence estimate is interpreted and incorporated in weighting schemes similarly to the heterogeneity estimate in a random-effects meta-analysis. To calculate the probability that an intervention is best on a series of cardiovascular outcomes, we used a Bayesian Markov chain Monte Carlo method.
Results: The combination of the different frequentist and Bayesian approaches allowed an understandable exploration of the consistency of the network and a summary of the evidence in terms of summary risk ratios, numbers needed to treat and probabilities that an intervention is best on a series of cardiovascular outcomes, which is understandable and meaningful for clinicians.
Conclusions: If several interventions are available for the same condition, a coherent analysis may only be achieved by summarising the entire network of trials comparing the different interventions while respecting randomization and carefully exploring the consistency of the different parts of the network.
Objectives: To present a combination of frequentist and Bayesian approaches towards exploring the consistency of the network and summarising direct and indirect comparisons in a single coherent and understandable analysis.
Methods: Using an example from interventional cardiology, we first conducted a standard random-effects meta-analysis separately for randomized comparisons of three pairs of interventions. Then, we used random-effects meta-regression to perform indirect comparisons of different interventions from trials that had one intervention in common. As a measure of inconsistency we calculated the I2 statistic. To combine direct within-trial comparisons with indirect between-trial comparisons, we used a frequentist approach to network meta-analysis, assuming two levels of random variation, heterogeneity and sampling error at the level of individual trials and incoherence at the level of pairs of interventions. The incoherence estimate is interpreted and incorporated in weighting schemes similarly to the heterogeneity estimate in a random-effects meta-analysis. To calculate the probability that an intervention is best on a series of cardiovascular outcomes, we used a Bayesian Markov chain Monte Carlo method.
Results: The combination of the different frequentist and Bayesian approaches allowed an understandable exploration of the consistency of the network and a summary of the evidence in terms of summary risk ratios, numbers needed to treat and probabilities that an intervention is best on a series of cardiovascular outcomes, which is understandable and meaningful for clinicians.
Conclusions: If several interventions are available for the same condition, a coherent analysis may only be achieved by summarising the entire network of trials comparing the different interventions while respecting randomization and carefully exploring the consistency of the different parts of the network.