執行網絡統合分析時標準化效果量之方法比較

Comparisons of statistical approaches to effect size standardization in network meta-analysis

陳韋翰(Wei-Han Chen)

指導教授 ： 杜裕康

國立臺灣大學/公共衛生學院/流行病學與預防醫學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103166

摘要

網絡統合分析（Network meta-analysis, NMA）是一種證據合成方法，藉由合併直接證據以及間接證據來同時比較多種治療。在進行網絡統合分析時，收錄的研究報告結果都是使用相同量表測量，會採用平均差異（mean difference, MD）作為報告之效果量。然而，在心理學或是教育學，收錄的研究報告之結果通常會使用不同量表來衡量。因此，在傳統的統合分析中為了確保不同量表測量的結果有可比較性，會將平均差異除以報告結果之標準差進行標準化，稱之為標準化效果量（Standardized mean difference, SMD）。現行的軟體套件計算混合標準差的方式是將該研究中所有報告的標準差混合在一起計算一個平均標準差，把傳統統合分析估計標準化效果量的方式延伸到網絡統合分析。這種作法背後隱藏的前提是收錄的研究不管是否使用相同量表，在同一篇研究中的所有治療效果皆是共享同一個標準差。
本篇的研究中我們先說明，假若選擇了不適當的標準化效果量的方法，會造成組內不一致性問題（within-trial inconsistency）以及組間不一致性問題（between-trial inconsistency）。我們接下來探討如何放寬現行方法中共享相同標準差的前提並提出新的contrast-based scale-pooling model以及arm regression model，結合研究使用何種量表的資訊或是考慮研究中不同治療之變異性來估計網絡統合分析中使用標準化效果量的結果（NMA-SMD）。後續的模擬顯示出本篇研究提出的模型與現行的方法相比，除了估計結果的效果量在所有模擬情境中不管是小樣本或是治療效果間的標準差是否相等，其估計誤差、最小均方差甚至是可接受型一誤差以及檢定力有較好的表現之外，也有同時解決網絡間組內不一致性問題及組間不一致性問題。
總結以上，在進行分析時，考慮到同時處理網絡間不一致性問題及估計結果的呈現時，當治療效果間標準差相等的情形下，應先採用本篇研究提出的contrast-based scale-pooling model，其估計結果相比於現行的方法更精準；當治療效果間標準差不等的情形下，則應先採用本篇研究提出的arm regression model，其估計的結果相比於現行的方法有較小的點估計誤差但有略大但可接受之最小均方差。

關鍵字

統合分析 ； 網絡統合分析 ； 標準化效果量 ； 組內不一致性 ； 組間不一致性 ； 實證醫學

並列摘要

Network meta-analysis (NMA) is an evidence synthesis tool that compares multiple treatments simultaneously by combining direct and indirect evidence. When the outcome of interest is continuous and measured with the same scale, mean difference is usually selected as the effect size for NMA. Yet, in psychological and educational research, the outcome is often measured with different scales in the included studies. To ensure comparability between different scales, standardized mean difference (SMD), which is obtained by dividing the mean difference by its standard deviation, is often the effect size of choice in pairwise meta-analyses. Current software packages extend SMD to NMA by pooling all the standard deviations of treatment effects within each study. This reflects the underlying assumption that treatment outcomes of the same study have a common standard deviation, which differs between studies, irrespective of the scales being used.
In this study, we highlight that within-trial and between-trial inconsistency can arise if the method of standardization is not appropriately chosen. We also seek to relax the assumption of shared standard deviation and incorporate information of common scales or consider treatment variability by proposing a novel contrast-based scale-pooling model and arm regression model for NMA using SMD as effect size (SMD-NMA). Extensive simulations showed that compared to current methods, not only the effect size estimates of our models had comparable bias and mean square error across all scenarios, and superior type I error rate and statistical power when either the sample size is small, or the standard deviation differs between treatments, but also resolved both within-trial inconsistency and between-trial inconsistency.
In conclusion, in addition to resolving the issues of inconsistency, our proposed contrast-based scale-pooling model is more powerful than current methods in homogeneous treatment variance and our arm regression model shows smaller biases than current methods but slightly larger mean square error in scenarios with heterogeneous treatment variances. Both of our models should be preferred methods for using SMD as the effect size in NMA.

並列關鍵字

meta-analysis ； network meta-analysis ； standardized mean difference ； within-trial inconsistency ； between-trial inconsistency ； evidence-based medicine

參考文獻

1. Beck, A.T., R.A. Steer, and M.G. Carbin, Psychometric properties of the Beck Depression Inventory: Twenty-five years of evaluation. Clinical psychology review, 1988. 8(1): p. 77-100.

2. Overall, J.E. and D.R. Gorham, The brief psychiatric rating scale. Psychological reports, 1962. 10(3): p. 799-812.

3. Jensen, M.P. and P. Karoly, Self-report scales and procedures for assessing pain in adults. 2011.

4. Cohen, J., Statistical power analysis for the behavioral sciences. Second ed. 1988: Hillsdale, N.J. : L. Erlbaum Associates.

5. Borenstein, M., et al., Introduction to meta-analysis. 2009: Chichester, U.K. : John Wiley Son.