Inconsistency between direct and indirect comparisons of competing interventions: meta-epidemiological study.
Song F., Xiong T., Parekh-Bhurke S., Loke YK., Sutton AJ., Eastwood AJ., Holland R., Chen Y-F., Glenny A-M., Deeks JJ., Altman DG.
OBJECTIVE: To investigate the agreement between direct and indirect comparisons of competing healthcare interventions. DESIGN: Meta-epidemiological study based on sample of meta-analyses of randomised controlled trials. Data sources Cochrane Database of Systematic Reviews and PubMed. Inclusion criteria Systematic reviews that provided sufficient data for both direct comparison and independent indirect comparisons of two interventions on the basis of a common comparator and in which the odds ratio could be used as the outcome statistic. MAIN OUTCOME MEASURE: Inconsistency measured by the difference in the log odds ratio between the direct and indirect methods. RESULTS: The study included 112 independent trial networks (including 1552 trials with 478,775 patients in total) that allowed both direct and indirect comparison of two interventions. Indirect comparison had already been explicitly done in only 13 of the 85 Cochrane reviews included. The inconsistency between the direct and indirect comparison was statistically significant in 16 cases (14%, 95% confidence interval 9% to 22%). The statistically significant inconsistency was associated with fewer trials, subjectively assessed outcomes, and statistically significant effects of treatment in either direct or indirect comparisons. Owing to considerable inconsistency, many (14/39) of the statistically significant effects by direct comparison became non-significant when the direct and indirect estimates were combined. CONCLUSIONS: Significant inconsistency between direct and indirect comparisons may be more prevalent than previously observed. Direct and indirect estimates should be combined in mixed treatment comparisons only after adequate assessment of the consistency of the evidence.