2020-08-012024-05-17https://scholars.lib.ntu.edu.tw/handle/123456789/673172摘要：統合分析整合文獻資料比較兩種治療之功效及危害。當治療方式大於兩種，雖可作 多次統合分析，但某些治療間可能無直接比較的隨機分派試驗、或有試驗但樣本數 不足以達成結論；各比較之結論亦可能不一致。 　　網絡統合分析整合直接與間接證據來克服上述限制。其方法學近來在模型比較 與不一致性方面多有進展。然而，以網絡統合分析行因果推論仍常相當困難。各試 驗之研究對象、治療方式及結果評估的異質性難以避免。若因而違背傳遞性假設 ，其對因果推論之影響評估亟須探討。 本計畫欲發展評估證據異質性造成網絡統合分析潛在偏誤之統計工具：透過研究偏 誤的網絡傳遞及對統計方法之影響，發展評估違反一致性和傳遞性之指標，並發展 測量訊息不確定性之指標。預定探討: 1.網絡結構如何決定偏誤及違反傳遞性之影響？偏誤之傳遞在各模型是否有差異 ，如何解釋？ 2.傳遞性存疑時應否調整估計之潛在影響？何種模型對違反傳遞性較為穩健？ 3.如何區分多治療試驗之直接和間接證據？區分偏誤對不一致性評估有何影響？ 4.如何評估網絡統合分析之不確定性？可否透過信賴區間判定？區間與排名不確定 性有何關係？ 5.現行排名不確定性之指標會受到治療數目影響，無法在不同分析間比較。新指標 的表現是否較佳？<br> Abstract: Traditional meta-analysis of interventions compares efficacy and harm of two treatment groups, but in most clinical scenarios, more than two treatments are usually available. If we wish to compare all available treatments, multiple pairwise meta-analyses may be undertaken, but this approach has several limitations. First, some treatments may not have been directly compared by any randomized controlled trials, or although they may have been compared, the results are inconclusive due to small numbers of patients involved. Secondly, conclusions from multiple pairwise meta-analyses may be inconsistent. Network meta-analysis was proposed to overcome these limitations by using both direct and indirect evidence. In recent years, substantial progress has been made about the methodology of network meta-analysis, especially in the issues of inconsistency and model comparison, but drawing causal inference from network meta-analysis remains a challenging task in many scenarios. Since no trials are identical, the heterogeneity in their patient populations, delivery of intervention and assessment out outcomes is inevitable. Currently, the research mainly focuses on whether or not the assumptions, such as consistency and transitivity, are violated, but a more important question remains unanswered, i.e. how to evaluate the potential impact of the violation of these assumptions on drawing causal inference? The objective of this 3-year project is to develop new statistical tools for evaluating the impact of potential biases due to heterogeneity in evidence on a network meta-analysis. We aim to measure quantitatively the consequence of violations of consistency and transitivity assumptions by investigating how biases propagate across a network and whether the impact of biases would be affected by the statistical approaches to network meta-analysis. Furthermore, we would like to develop statistical indices to measure the uncertainty of information within a network meta-analysis. 1.How is the impact of biases and intransitivity determined by network geometry? Are the routes of bias propagation different in the Lu & Ades model, baseline model and arm-based model? How to explain the differences in the routes of bias propagation? 2.When transitivity assumption is questionable, should adjustment be undertaken to address its potential impact on the estimates of a network meta-analysis? Which statistical models, namely the Lu & Ades model, baseline model and armbased model are more robust to the violation of transitivity assumption? 3.How to separate the direct evidence from the indirect evidence in multi-armed trials adequately? How does inadequate separation of direct from indirect evidence affect our evaluation of the consistency assumption? 4.How to evaluate the degree of uncertainty of each network meta-analysis? Should the confidence interval of each pairwise comparison be used to determine the uncertainty of evidence? What is the relation between the uncertainty of ranking and the confidence interval of each pairwise comparison? 5.Some authors claimed that 95% Credible Interval of SUCRA is an uninformative index, as its range is restricted to the number of treatments. Consequently, the credible intervals are not comparable across network metaanalyses. Are our proposed two indices better tools for measuring the uncertainty of the ranking of treatments?統合分析網絡統合分析一致性傳遞性熵meta-analysisnetwork meta-analysisconsistencytransitivityentropy