Article type
Year
Abstract
Background: In the absence of head-to-head randomized controlled trials (RCTs), indirect treatment comparisons (ITC) and network meta-analysis can be performed to inform healthcare decision-making. Mixed treatment comparisons (MTC), a special case of network meta-analysis, consider a greater share of the evidence base by combining direct with indirect comparisons. It is known that network meta-analysis of RCTs result in invalid estimates when there are differences in treatment effect modifiers across comparisons.
Objective: To explain heuristically when network meta-analyses result in invalid findings.
Methods: By means of directed acyclic graphs (DAGs), which are commonly used in epidemiology to guide analyses, it is illustrated when adjustment for covariates in a network meta-analysis of RCTs is required to minimize bias, when adjustment is unnecessary, and when adjustment for covariates can in fact introduce bias.
Results: Although statistically adjusting for differences in modifiers of the relative treatment effect across comparisons can remove bias in an ITC or MTC, adjustment for differences that are not effect modifiers is not necessary and can introduce collider stratification bias. As a special case we argue that adjustment for differences in discontinuation rates across comparisons might be necessary, but can introduce bias when discontinuation is an effect of the outcome. Furthermore, the adjustment for the baseline risk in a network meta-analysis might be useful to explain heterogeneity, but can also introduce bias.
Conclusion: When meta-regression models or subgroup analysis are used in a network of interventions in an attempt to improve comparability of trials, one has to be aware that adjustment for some covariates can in fact introduce bias. DAGs can be considered useful to guide ITC and MTC analyses to minimize the risk of biased estimates.
Objective: To explain heuristically when network meta-analyses result in invalid findings.
Methods: By means of directed acyclic graphs (DAGs), which are commonly used in epidemiology to guide analyses, it is illustrated when adjustment for covariates in a network meta-analysis of RCTs is required to minimize bias, when adjustment is unnecessary, and when adjustment for covariates can in fact introduce bias.
Results: Although statistically adjusting for differences in modifiers of the relative treatment effect across comparisons can remove bias in an ITC or MTC, adjustment for differences that are not effect modifiers is not necessary and can introduce collider stratification bias. As a special case we argue that adjustment for differences in discontinuation rates across comparisons might be necessary, but can introduce bias when discontinuation is an effect of the outcome. Furthermore, the adjustment for the baseline risk in a network meta-analysis might be useful to explain heterogeneity, but can also introduce bias.
Conclusion: When meta-regression models or subgroup analysis are used in a network of interventions in an attempt to improve comparability of trials, one has to be aware that adjustment for some covariates can in fact introduce bias. DAGs can be considered useful to guide ITC and MTC analyses to minimize the risk of biased estimates.