網絡統合分析中偏差的傳導模式

Bias propagation in network meta-analysis models

李驊(Hua Li)

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(2018年)

https://doi.org/10.6342/NTU201802673

摘要

研究目的
研究偏差的治療效果，在不同網絡結構與不同的模型中會如何影響估計值的結果。
研究方法
以頻率學派的架構執行Lu Ades model、Baseline model和Arm-based model。Arm-based model的相關係數矩陣設定為治療彼此之間互相獨立，每個治療允許有各自的隨機效應。三種模型皆以二階段分析處理。第一階段先以固定效應模型估計，以其結果估計隨機效應變異數的值。第二階段將前面估計出的變異數作為隨機效應模型的權重參數，再估計(相對或絕對)治療效果。Arm-based model的絕對治療效果將被轉換成相對治療效果後，與其他兩個模型做比較。透過模擬不同的網絡結構下七種偏差情境，來探討偏差在不同模型下傳遞的模式。
研究結果
Lu Ades model執行前需將治療做排序，一組成對比較中不論偏差的治療是哪一個，偏差的影響會往兩個治療中排序靠後的治療的估計值傳遞。當兩個治療偏差的方向與大小相同時，偏差影響會互相抵消使估計結果不受影響。偏差在Baseline model裡主要影響偏差的治療的相對效果估計值，因此偏差的影響較能如實呈現在對應治療的估計值上。但部分偏差的影響會透過基底治療傳遞到整個網絡，這部分對治療效果評估的影響類似Lu Ades model，傾向傳遞給排序靠後的治療的估計值。Arm-based model的偏差會如實地在偏差的治療的相對效果估計值上觀察到，其他完全不會受到影響。
結論
三種模型的偏差情況各有特色，區別在於要由哪些估計值去吸收偏差的效果，因此不存在最好的模型。研究者可以根據資料有疑慮、可能存在偏差的部分，評估使用何種模型最有利，或是較能預期估計值偏差的大小與方向。

關鍵字

網絡統合分析 ； contrast-based model ； arm-based model ； 二階段模式 ； 偏差

並列摘要

Aims: To investigate how bias contained in direct evidence of one treatment contrast within a network meta-analysis will impact on the estimation of other treatment contrasts under the Lu Ades model, baseline model, and arm-based model.

Methods: Simulations of the three statistical models for network meta-analysis within the frequentist framework were undertaken to evaluate the impact of bias in evidence on one treatment contrast. The within-study correlation structure in the arm-based model was assumed to be independent and heterogeneous. The analyses of the three models used two-step methods by first pooling together all pairwise comparisons for each treatment contrast and then undertaking a fixed effect network meta-analysis. This result was used to calculate a variance by the sum square of Pearson residual, and then in the second step we estimated the relative or absolute treatment effect with the variance becoming weighted parameter in random effect model. We simulated seven scenarios to evaluate how different geometrical structures of a network affect the impact of the bias in the evidence of a treatment contrast on the estimates of other treatment contrasts.

Results: In Lu Ades model, treatments were arranged in an order that started with the global baseline treatment. The bias in one treatment affected the estimate of treatments that were farther in the order. When the biases in the two treatments of a contrast had the same direction and magnitude, the estimated bias in this treatment contrast was cancelled out. In the baseline model, the bias in one treatments partially affected the estimate of this treatment, while some of the bias was propagated through the baseline treatment, yielding small bias in the other relative effects. Under the arm-based model, the effect of the biases did not spread to other treatment contrasts unrelated to the biased treatment.

Conclusions: Bias propagation in three different models has various patterns which are also affected by the geometry of the network. When evidence of some treatments may contain biases, researchers could then evaluate which estimates may be affected by these biases and how much the potential impact is.

並列關鍵字

network meta-analysis ； contrast-based model ； arm-based model ； two-stage model ； bias

參考文獻

參考文獻

1. Guyatt, G., Cairns, J., Churchill, D., and et al. (1992). Evidence-based medicine: A new approach to teaching the practice of medicine. JAMA 268, 2420-2425.

2. Uman, L.S. (2011). Systematic Reviews and Meta-Analyses. Journal of the Canadian Academy of Child and Adolescent Psychiatry 20, 57-59.

3. Efthimiou, O., Debray, T.P., van Valkenhoef, G., Trelle, S., Panayidou, K., Moons, K.G., Reitsma, J.B., Shang, A., and Salanti, G. (2016). GetReal in network meta-analysis: a review of the methodology. Res Synth Methods 7, 236-263.

4. Biondi-Zoccai, G. (2016). Umbrella Reviews: Evidence Synthesis with Overviews of Reviews and Meta-Epidemiologic Studies.(Springer International Publishing).