Article type
Abstract
Background: Network meta-analysis (NMA) can be performed either under a frequentist (classical) or a Bayesian framework. With recent developments in frequentist software, more researchers use this approach for NMA; however, the extent to which the results of these approaches yield similar results remains uncertain.
Objectives: Our goal was to investigate the variability in results from frequentist and Bayesian approaches comparing the direct, indirect, and mixed-effect estimates as well as the ranking of the interventions in a sample of published networks.
Methods: We performed a systematic survey of the literature and included a sample of systematic reviews of randomised controlled trials (RCT) from the field of cardiovascular medicine that used NMA methods to compare the effects of more than two interventions with a dichotomous primary outcome. Eligible studies have to provide enough data to re-run the analysis including interventions assessed in each trial, number of events and number of patients per arm. To perform frequentist NMA network suite commands, STATA version 14.1 was used. The gemtc package (version 0.8, released on 2016-03-01) in R software was used for vague prior-Bayesian NMA.
Results: We re-analysed data from 14 NMAs. Included NMAs had 12 to 63 RCTs informing 4 to 12 interventions. On average, the absolute difference between Bayesian and frequentist odds ratios were 0.18 ± 0.20 across all comparisons (range from 0.00 to 0.65) in a fixed-effects model. For a random-effects model, the average absolute difference between Bayesian and frequentist odds ratios were 0.26 ± 0.44 across all comparisons (range from 0.00 to 1.58). Node-splitting results were almost similar in both approaches. SUCRA values were slightly different between the two approaches but most of the time treatment rankings were the same.
Conclusions: Our findings showed that magnitude of the effect estimates, but rarely the direction or treatment rankings, may differ to a large extent between Bayesian and frequentist approaches.
Objectives: Our goal was to investigate the variability in results from frequentist and Bayesian approaches comparing the direct, indirect, and mixed-effect estimates as well as the ranking of the interventions in a sample of published networks.
Methods: We performed a systematic survey of the literature and included a sample of systematic reviews of randomised controlled trials (RCT) from the field of cardiovascular medicine that used NMA methods to compare the effects of more than two interventions with a dichotomous primary outcome. Eligible studies have to provide enough data to re-run the analysis including interventions assessed in each trial, number of events and number of patients per arm. To perform frequentist NMA network suite commands, STATA version 14.1 was used. The gemtc package (version 0.8, released on 2016-03-01) in R software was used for vague prior-Bayesian NMA.
Results: We re-analysed data from 14 NMAs. Included NMAs had 12 to 63 RCTs informing 4 to 12 interventions. On average, the absolute difference between Bayesian and frequentist odds ratios were 0.18 ± 0.20 across all comparisons (range from 0.00 to 0.65) in a fixed-effects model. For a random-effects model, the average absolute difference between Bayesian and frequentist odds ratios were 0.26 ± 0.44 across all comparisons (range from 0.00 to 1.58). Node-splitting results were almost similar in both approaches. SUCRA values were slightly different between the two approaches but most of the time treatment rankings were the same.
Conclusions: Our findings showed that magnitude of the effect estimates, but rarely the direction or treatment rankings, may differ to a large extent between Bayesian and frequentist approaches.