Article type
Year
Abstract
Background: Techniques to assess model validity in network meta-analyses include: convergence, model fit and consistency analysis. A systematic review and network meta-analysis of acupuncture and other physical therapies for the relief of chronic pain due to osteoarthritis of the knee (OAK) was undertaken.
Objectives: The impact of the different quality inclusion criteria on these three aspects of model validity in the context of this systematic review are discussed.
Methods: Studies were divided up into whether or not they were at least good quality or were poor quality according to a CRD quality checklist. In a sensitivity analysis, extreme outliers were excluded from the analyses. Each of these studies were of poor quality so this did not change the networks excluding poor quality studies.
Results: Convergence was achieved in each of these analyses even though model fit was not perfect for the networks including poor quality studies. Model fit was far better for the networks excluding poor quality studies. There was considerably more evidence of inconsistency in networks including poor quality studies. There were, however, far fewer triangles of evidence in the networks excluding poor quality studies and therefore little potential for inconsistency. The elimination of the extreme outliers from the analyses improved model fit significantly, suggesting that it was within-comparison heterogeneity rather than the presence of inconsistency that caused the poor model fit in this case. The outliers were not responsible for the inconsistency.
Conclusions: The presence of inconsistency suggests that the uncertainty around the estimates has been underestimated. The absence of evidence of inconsistency when it is not possible to test for it does not mean heterogeneity has been adequately accounted for. In this case, outliers and within-comparison heterogeneity were responsible for poor model fit, but not the inconsistency.
Objectives: The impact of the different quality inclusion criteria on these three aspects of model validity in the context of this systematic review are discussed.
Methods: Studies were divided up into whether or not they were at least good quality or were poor quality according to a CRD quality checklist. In a sensitivity analysis, extreme outliers were excluded from the analyses. Each of these studies were of poor quality so this did not change the networks excluding poor quality studies.
Results: Convergence was achieved in each of these analyses even though model fit was not perfect for the networks including poor quality studies. Model fit was far better for the networks excluding poor quality studies. There was considerably more evidence of inconsistency in networks including poor quality studies. There were, however, far fewer triangles of evidence in the networks excluding poor quality studies and therefore little potential for inconsistency. The elimination of the extreme outliers from the analyses improved model fit significantly, suggesting that it was within-comparison heterogeneity rather than the presence of inconsistency that caused the poor model fit in this case. The outliers were not responsible for the inconsistency.
Conclusions: The presence of inconsistency suggests that the uncertainty around the estimates has been underestimated. The absence of evidence of inconsistency when it is not possible to test for it does not mean heterogeneity has been adequately accounted for. In this case, outliers and within-comparison heterogeneity were responsible for poor model fit, but not the inconsistency.