忽略元件交互作用對元件網絡統合分析中治療效應之估計偏誤

Abstract

在比較多種治療時，可以使用網絡統合分析 (Network Meta-Analysis, NMA) 來估計不同治療之間的相對效果，一般來說，Lu & Ades模型較廣為使用，雖然能結合直接證據與間接證據來比較多種治療，但當治療間缺乏直接或間接比較時，此時，就無法直接使用NMA來處理斷裂的網絡。 元件網絡統合分析 (Component Network Meta-Analysis, CNMA) 中，透過相加性 (additivity) 的假設將治療視為由多個元件組成，並假設治療效果是各元件效果相加來解決斷裂網絡的問題。然而，元件間可能存在交互作用，協同與拮抗作用使得治療效果可能較好或較差，若在使用CNMA模型時，忽略元件間的交互作用，則可能導致治療效果估計的偏誤。本研究是基於CNMA模型，探討在已知元件間存在交互作用的前提下，但在使用CNMA模型時忽略了元件間交互作用後，對各元件治療效應估計的偏誤影響。 研究先設計了不同情境，探討在元件間存在交互作用時，CNMA模型中各元件效果的估計偏誤。通過數學推導和矩陣運算，確定了元件效果偏誤的估計式，並發現在特定情況下，元件偏誤估計值會與網絡結構無關，而主要受到納入到元件網絡統合分析的是哪些治療種類的影響，亦即不管網絡結構如何改變，只要網絡圖是由相同種類的治療所構成，那麼各元件的偏誤估計值在不同網絡結構下仍然相同。 最後，為了更直觀地展示元件偏誤，研究使用了視覺化方法，將元件視為節點，連線代表元件組合的治療，並於圖中呈現各元件的偏誤大小。未來研究應繼續探討多個交互作用的影響，並改進視覺化方法，提供更直觀的圖形呈現各元件的偏誤。

When comparing multiple treatments, a network meta-analysis (NMA) model can be used to estimate the relative effects between different treatments. Generally, the Lu & Ades model is widely used. It can combine direct and indirect evidence to compare multiple treatments. However, when there is a lack of direct or indirect comparisons between treatments, NMA cannot be used directly to address disconnected networks . Component Network Meta-Analysis (CNMA) addresses the issue of disconnected networks by assuming additivity, viewing treatments as composed of multiple components, and assuming that the treatment effect is the sum of the effects of each component. However, interactions between components, such as synergistic or antagonistic effects, can make the treatment effect better or worse. Ignoring these interactions when using the CNMA model can lead to biased estimates of treatment effects. Hence, this study investigates the impact of such bias on the estimation of each component's treatment effect under the CNMA model when known interactions between components are ignored. The study first designed several scenarios to explore the bias of each component effect in the CNMA model when an interaction between components are present. Through mathematical derivation, it is determined that the estimation bias of the component effects is actually independent of the network structure and is mainly influenced by the types of treatments included in the CNMA under specific conditions. Hence, the bias of each component effect in the CNMA model would affected by the network structure as well as the treatments included in the analysis. Finally, to more intuitively display the component biases, the study uses visualization methods, treating components as nodes with the label of each component's bias shown in the plot. Future research should continue to explore the effects of multiple interactions and improve visualization methods to provide a more intuitive graphical presentation of each component's bias.