- 學位論文
利用結構方程模式進行診斷工具表現比較的統合分析
Applying structural equation modeling to meta-analysis of diagnostic test performance
摘要
背景：受限於敏感性與特異度在研究之內與之間的相關性，診斷型網絡統合分析的方法較少被提出，且目前並沒有一個公認的標準模型，也沒有文獻在統一的框架下，完整地比較現今已發表的方法的不同之處，而結構方程模式是一個非常靈活的統計框架，其能夠使用潛在變量將隨機效應明確地建模，因此能夠輕易地指定敏感性和特異度之間潛在的相關性，此外，藉由路徑圖的輔助，還能直觀地比較模型之間的異同，因此，本研究之目的在於探討如何在結構方程模式的框架中，對診斷型網絡統合分析建模，並比較模型之間的異同，以及參數估計表現的差異。 方法：本研究使用結構方程模式的架構，建立四種基於臂的診斷型研究之網絡統合分析模型，其中兩種為由Trikalinos以及Dimou分別提出，有考慮到研究內相關性的方法；而另外兩種為由Hoyer和Nyaga分別提出，沒有考量這層關係的方法。我們使用路徑圖視覺化以及指定這四個模型，藉此直觀比較模型之間的差異。之後藉由模擬研究的方式，針對評估兩種診斷工具表現的情況，我們創造了九個不同的情境，其中使用八種評估指標為標準，如Bias、EmpSE、MSE以及Coverage等等，來評估這四種模型在估計診斷工具敏感性和特異度的表現。 結果：由路徑圖可以直觀地看出，Nyaga可以被視為Hoyer的簡化版，而Dimou則可理解為Hoyer的複雜版，至於Trikalinos則比Hoyer多使用了JTPR與JFPR來捕捉組內相關性。另外模擬研究的結果顯示，Trikalinos在診斷工具之間無相關的情況下，收斂率很低，除此之外，四個模型在九個情境下，參數估計的EmpSE與MSE表現並沒有相差很多，但就其他評估指標來看，有考慮到組內的相關性的Trikalinos和Dimou，在高相關或敏感性和特異度參數真值設定接近1時估計會較為不穩定，且Bias普遍較Hoyer和Nyaga高。 結論：Trikalinos或Dimou的參數估計易受到高相關以及真值設定的影響，因此，應注意真實參數的情況以及診斷工具之間的相關性為何，方能得到較精確的估計，而Hoyer或Nyaga兩個模型相差不多，且整體而言，此兩個模型的準確度會比Trikalinos或Dimou好一點。而在模型選擇的部分，除了考量上述的估計表現外，還應考慮另外三個層面，一為可用的資料型態為何；二為是否應考慮組內相關性；三為欲比較的診斷工具數目為何。
並列摘要
Background: Because of the inherent correlation between sensitivity and specificity, evidence synthesis of diagnostic test accuracy data is more complicated than standard meta-analysis. Statistical methods for evaluating the performance of multiple diagnostic tests have emerged only recently. Currently no statistical approach has been widely considered as the standard one, so a comprehensive evaluation of the proposed approaches in the literature, including an assessment of model performance under different criteria and scenarios, would be a valuable addition to the current literature. Previous studies have shown that network meta‐analysis (NMA) can be implemented in the structural equation modeling (SEM) framework. SEM is a flexible statistical framework as random effect is explicitly modeled as a latent variable; hence the SEM-based NMA on diagnostic accuracy studies can specify the inherent correlation between sensitivity and specificity explicitly. Therefore, the aim of this study is to show how to undertake a network meta-analysis of diagnostic test within the statistical framework of SEM, and to compare several existing approaches under the unified framework. Methods: We implemented 4 different types of NMA models on diagnostic accuracy studies into the SEM framework. Two of them are proposed by Dimou and Trikalinos, respectively, which account for potential within-studies correlations, and the other two are proposed by Hoyer and Nyaga, respectively, which do not explicitly consider these correlations. Then, the path diagram was used to specify these 4 models in order to visually compare the differences between the models. Furthermore, we conducted a simulation study for the comparison of two diagnostic tests with the gold standard to assess the performance of the parameters estimates in these models. In total, 9 scenarios and 8 performance indices (such as bias, EmpSE, MSE, coverage, etc.) were used in the simulation to evaluate the model performance in terms of the estimation of sensitivity and specificity. Results: It can be seen intuitively from the path diagram that Nyaga’s model is a simplified version of Hoyer, and Dimou’s model is a complex version of Hoyer’s. Besides, Trikalinos’s model is also similar to Hoyer’s but it accounts for the within-studies correlations by modeling the logit-JTPR and logit-JFPR. However, simulations suggest that Trikalinos's model often failed to converge, when the correlation between the diagnostic tests is null or low. In other situations, the indices of EmpSE and MSE of parameters estimates in each model showed similar values. However, after considering other performance measures, if the within-studies correlation is set to high or the parameters of sensitivity and specificity are close to 1, the estimators of Trikalinos and Dimou will become unstable. In general, the biases in Trikalinos’s and Dimou’s models are larger than those of Hoyer’s and Nyaga’s. Conclusions: The parameter estimates of Trikalinos’s or Dimou’s models are susceptible to high within-study correlations and high sensitivity/specificity. On the contrary, the estimates given by Hoyer’s and Nyaga’s models are relatively stable, and results from these two models are quite consistent and almost identical. The accuracy of Hoyer’s and Nyaga’s models are better than those of Dimou and Trikalinos.