A Leave-One-Out Algorithm for Contribution Analysis in Additive Component Network Meta-Analysis

Background Component network meta-analysis (CNMA) enables disentangling individual treatment component effects but faces challenges in quantifying evidence contributions due to path enumeration complexity and component pathway unidentifiability. This study aims to develop a novel leave-one-out algorithm to address these limitations and enhance the interpretability of contribution metrics for individual component effects. Methods We propose a precision-centric leave-one-out algorithm grounded in jackknife principles. By iteratively removing direct comparisons and measuring inverse variance perturbations, the method quantifies statistical leverage via a contribution matrix. Parameter decomposition separates direct evidence from additive network evidence, ensuring component effects are estimated without contamination while independently weighting direct evidence. In the absence of prior benchmarks, we employed a surrogating internal validation through linear weighting to validate the congruence between contribution-weighted predictions and model-derived estimates. Results Application across illustrative scenarios demonstrated the algorithm’s utility in identifying critical comparisons that stabilize component estimates. Validation on a real dataset (21 components, 40 interventions, 66 comparisons) revealed high congruence between contribution-weighted predictions and additive model estimates: Pearson correlation r = 0.953 (p < 0.001), explained variance R² = 0.907, and mean absolute error MAE = 0.172. Coherence tests showed minimal discrepancies (< 0.1%) between combined estimates (integrating direct and additive evidence) and full model estimates. Conclusions The leave-one-out algorithm redefines contribution analysis in additive CNMA by replacing path enumeration with precision leverage quantification. It resolves unidentifiability challenges, enhances evidence transparency, and supports optimization of multi-component interventions. Limitations include computational scalability and heuristic redistribution rules, warranting future refinements.