Sensitivity of treatment recommendations to bias in NMA (Phillippo threshold framework)
This method helps assess how strong the evidence is in a network meta-analysis by showing how much bias would need to exist to change the recommended treatment. It calculates thresholds that indicate how sensitive a decision is to potential biases in the data.
At a glance
Use when
Conducting sensitivity analysis in network meta-analysis to assess decision robustness; when risk of bias may influence treatment rankings; to support decision-making under uncertainty
Avoid when
The NMA model structure or covariance matrix is unavailable; when bias is expected to act non-additively or through complex mechanisms not captured by simple adjustments
Inputs
Network meta-analysis data, including treatment effects and variance-covariance structure; optionally, risk of bias assessments for individual studies
Outputs
Bias adjustment thresholds indicating the minimal changes needed to alter treatment recommendations; sensitivity measures for individual studies or treatment contrasts
How it works
The method computes bias adjustment thresholds using efficient matrix operations to determine the minimal changes in study data that would alter treatment recommendations in fixed or random-effects network meta-analysis models. It supports exploration of bias impacts at the level of individual studies or aggregate treatment contrasts. For complex models with multiple data types, an approximation to the hypothetical aggregate likelihood is used. Implemented in an R package.
- HTA domains
- Clinical Effectiveness, Costs & Economic Evaluation
- Assumptions
- The impact of bias can be approximated by adjusting study-level effect estimates; the NMA model is correctly specified; the decision is based on relative treatment effects
- Strengths
- Provides quantitative, interpretable thresholds for decision robustness; applicable to both fixed and random-effects models; handles complex NMAs via likelihood approximation; enables targeted sensitivity analysis
- Limitations
- Relies on approximations for complex data types; assumes bias acts additively on effect estimates; does not model bias mechanisms directly; requires access to NMA model structure and covariance matrix
- Also known as
- Phillippo threshold framework, Bias adjustment thresholds in NMA
Questions this answers
- › How sensitive is a treatment recommendation to potential bias in a network meta-analysis?
- › Which studies or treatment comparisons most influence the robustness of the decision?
- › How much bias would need to be present to change the preferred treatment?
- › What is the impact of risk of bias on decision-making in NMA?
- › Can we quantify the robustness of NMA-based decisions to data perturbations?
- › How can we explore the consequences of bias in both simple and complex NMA models?
References & sources
Similar by meaning
Beta record. Generated from the primary source via AI extraction and independent audit, pending final human review.