Article type
Year
Abstract
Background: Selection models are used to explore the potential impact of publication bias (PB) via sensitivity analysis by making assumptions for the probability of publication of trials conditional on their precision. Selection models have been previously extended for a star-shaped network where several treatments are compared to a common comparator but not between themselves.
Objectives: (a) to suggest a selection model for PB in a full Network Meta-Analysis (NMA) (b) to explore whether different trial designs (number of arms and nature of comparison) pertain to different levels of PB. More specifically, we explore whether multi-arm and head-to-head trials are less prone to PB than two-arm and placebo-controlled trials.
Methods: We developed a design-by-PB interaction model to describe the mechanism by which studies with different designs and precision are selected for publication. We measure the extent of PB by the correlation coefficient between propensity for publication and study results. We illustrate the methodology in a network including two-arm and three-arm trials that compare Placebo, Aspirin and Aspirin plus Dypiridamole for the failure of vascular graft or arterial patency.
Results: The correlation between probability of publication and effect size for the comparison Aspirin versus Placebo is larger in two-arm studies compared to three-arm studies. In our example, publication of a three-arm study is not dependent on the estimated relative treatment effects, unlike placebo-controlled trials for which publication is highly associated with the magnitude of the treatment effect.
Conclusion: The suggested selection model accounts for the fact that larger studies, studies with more than two arms and head-to head studies might have larger chances for publication independently of their findings. We suggest employing this sensitivity analysis across various scenarios for PB to infer about the robustness of the summary treatment effects and the ranking of the competing treatments.
Objectives: (a) to suggest a selection model for PB in a full Network Meta-Analysis (NMA) (b) to explore whether different trial designs (number of arms and nature of comparison) pertain to different levels of PB. More specifically, we explore whether multi-arm and head-to-head trials are less prone to PB than two-arm and placebo-controlled trials.
Methods: We developed a design-by-PB interaction model to describe the mechanism by which studies with different designs and precision are selected for publication. We measure the extent of PB by the correlation coefficient between propensity for publication and study results. We illustrate the methodology in a network including two-arm and three-arm trials that compare Placebo, Aspirin and Aspirin plus Dypiridamole for the failure of vascular graft or arterial patency.
Results: The correlation between probability of publication and effect size for the comparison Aspirin versus Placebo is larger in two-arm studies compared to three-arm studies. In our example, publication of a three-arm study is not dependent on the estimated relative treatment effects, unlike placebo-controlled trials for which publication is highly associated with the magnitude of the treatment effect.
Conclusion: The suggested selection model accounts for the fact that larger studies, studies with more than two arms and head-to head studies might have larger chances for publication independently of their findings. We suggest employing this sensitivity analysis across various scenarios for PB to infer about the robustness of the summary treatment effects and the ranking of the competing treatments.