A leave-one-out algorithm for contribution analysis in component network meta-analysis.

backgroundComponent network meta-analysis (CNMA) enables disentangling individual component effects from multicomponent treatments. However, no established methods exist to quantify the contribution of evidence from constituent comparisons to the disentangled component effect estimates in CNMA, hindering the interpretability of results.

methodsWe proposed a leave-one-out algorithm to address this gap. The core approach iteratively excludes each constituent comparison (i.e., edge in the network), recomputes the variances of all component effects, and quantifies the precision leverage of each comparison based on the induced variance inflation. Contributions are assigned via a normalized matrix. We developed special rules to handle cases where exclusion renders component effects unidentifiable. The method also formally decomposes component estimates into direct and additive evidence sources. Its utility and validity were evaluated through implementation using hypothetical networks and a real-world dataset.

resultsThe leave-one-out algorithm accurately identified pivotal evidence sources by capturing substantial variance fluctuations upon their exclusion. Contributions assigned via precision leverage effectively quantified the critical importance of comparisons isolating target components. Application to real-world data (66 comparisons, 21 components) also confirmed the method's precision in discerning influential evidence within complex networks, and exhibited strong alignment with the parameter decomposition results. Crucially, validation revealed no inherent relationship exists between precision leverage and linear weighting.

conclusionsThe leave-one-out algorithm resolves a critical gap in CNMA methodology by providing a robust, variance-based framework for quantifying the contribution of constituent direct comparisons to component effect estimates. It reliably identifies pivotal evidence sources essential for component identifiability and precision across diverse network structures, enhancing the transparency and interpretability of evidence synthesis for complex interventions.