网状Meta分析：读者、作者和审稿人的方法论点

分享智慧

共同成长

Full text

Network meta-analysis (NMA) is a useful statistical method that allows comparison of multiple treatments to be considered in a single analysis by combining direct with indirect evidence. The BJD has seen an increase in submissions of systematic reviews employing NMA over the past couple of years;1 therefore, we now provide methodological guidance to help authors submit a high-quality NMA.

Direct evidence is often obtained from randomized controlled trials while indirect evidence can be mathematically deduced when two or more interventions have been compared with a common comparator. For example, in a recent Cochrane Database systematic review and NMA, 20 systemic treatments for moderate-to-severe psoriasis were considered.2 The relative effect of infliximab vs. secukinumab – for which no study is available – was estimated indirectly via comparisons with placebo (Figure 1). NMA also allows one to rank treatments, thus answering an important question for physicians, patients and guideline authors: among all available treatments, which works best?

Given the growing spike in publications related to NMA, concerns have emerged regarding their methodological quality.3 To ensure validity of findings, it is fundamental that authors accurately plan, conduct and report a NMA. This includes the formulation of a precise, clinically pertinent research question, the conduct of a thorough systematic review, assessment of the assumptions of NMA, transparency and comprehensive presentation of results, and the evaluation of risk of bias and certainty of the evidence. A protocol, which outlines these stages, needs to be prospectively registered. Authors should follow the PRISMA extension statement for systematic reviews with NMA to ensure comprehensiveness and transparency of reporting.4

A well-formulated question is crucial in guiding authors throughout the NMA, from the definition of eligibility criteria to the reporting of findings, and will help to determine which populations and treatments to include in the network, and thus the shape the network of evidence may take. Decisions of whether the different interventions should be evaluated as individual drugs, specific doses, or lumped into drug classes need to be made in consideration of the research question and the underlying assumptions, notably the assumption of transitivity.5

Transitivity refers to the validity of carrying out indirect comparisons via an intermediate treatment and is a fundamental assumption of NMAs. It assumes there are no major differences between the included studies regarding all important factors that may affect the outcome, such as patient characteristics and disease severity. For example, trials involving co-interventions and biological-naïve participants were excluded from a systematic review as they would have induced intransitivity.2 Therefore, authors should consider, for example, the eligibility of trials of co-interventions that are known to be associated with higher efficacy compared with monotherapy.

Discrepancies in the distributions of effect modifiers manifest in the data as disagreement between direct and indirect estimates, known as statistical incoherence, and can sometimes also be a source of important heterogeneity. Several statistical tests exist and should be used to check coherence, both globally (in the whole network) and locally (in parts of the network). If incoherence and/or heterogeneity is present, subgroup analyses and network meta-regression may be used to further identify the potential sources.

Another key step is the evaluation of publication bias, where assessment of small-study effects constitutes an important step. This is checked visually through a modified version of the meta-analysis funnel plot called ‘comparison-adjusted funnel plot’. When large asymmetries are present in the plot, small-study effects are likely acting. Network meta-regression can help to identify the causes. Additionally, sensitivity analyses should always be planned and conducted to assess if the results are robust to different methodological choices, such as the exclusion of small studies or studies at high risk of bias.

A clear presentation of the findings is paramount and can be challenging to produce, especially when the network is large. The overall network effects are usually reported in forest plots while the relative effects between every combination of treatments are summarized in league tables. When many treatments are available, the number of two-by-two relative effects quickly becomes very large: for example, in the network of 20 treatments, the number of two-by-two comparisons reached 190.2 An advantage of NMA is its ability to provide a coherent ranking of treatments, for which the most popular metric is the Surface Under the Cumulative Ranking Curve (SUCRA).6 SUCRA values range between 0% and 100% (the higher the value, the higher the likelihood that the treatment is top ranked). However, it is important for these to be interpreted in conjunction with the relative effects results otherwise misleading conclusions can be made. For example, the SUCRA values of the Psoriasis Area and Severity Index 90 outcome for infliximab, secukinumab and brodalumab were 93·6, 76·2 and 68·4, respectively,2 but when comparing the two-by-two relative effects with each other, these three drugs did not show significant statistical differences in efficacy due to large uncertainty in the results. Thus, ranking measures should always be reported with the relative effects.

Furthermore, several approaches have been developed to evaluate the certainty of the evidence obtained from NMAs.7, 8 CINeMA (Confidence In Network Meta-Analysis: http://cinema.ispm.ch/) is a web application extending GRADE (Grading of Recommendations Assessment, Development and Evaluation) that considers six domains to evaluate the certainty of the evidence: within-study bias, reporting bias, indirectness, imprecision, heterogeneity and incoherence. Rating the certainty of evidence with these approaches enhances the transparency, reproducibility and credibility of the results.

In summary, NMAs are complex and challenging but if well conducted, they can provide the highest level of evidence in comparative effectiveness research. There is a need for collaborative work when conducting NMAs between expert clinicians, those with expertise in the conduct of systematic reviews, and methodologists and statisticians experienced in NMA. International efforts are needed to encourage authors and reviewers to follow the existing guidelines to limit the publication of poor-quality NMAs.

全文翻译（仅供参考）

网络荟萃分析 (NMA) 是一种有用的统计方法，通过将直接证据与间接证据相结合，可以在单个分析中考虑多种治疗方法的比较。在过去的几年中， BJD提交的使用 NMA 的系统评价有所增加；1因此，我们现在提供方法指导来帮助作者提交高质量的 NMA。

直接证据通常从随机对照试验中获得，而间接证据可以在将两种或多种干预措施与一个共同的比较器进行比较时从数学上推导出来。例如，在最近的Cochrane 数据库系统评价和 NMA 中，考虑了 20 种中度至重度银屑病的全身治疗。2英夫利昔单抗与苏金单抗的相对效应（尚无研究可用）是通过与安慰剂的比较间接估计的（图 1）。NMA 还允许对治疗进行排名，从而回答医生、患者和指南作者的一个重要问题：在所有可用的治疗中，哪种治疗效果最好？

鉴于与 NMA 相关的出版物数量不断增加，人们对其方法学质量表示担忧。3为确保研究结果的有效性，作者准确计划、实施和报告 NMA 至关重要。这包括制定精确的、与临床相关的研究问题、进行彻底的系统评价、评估 NMA 的假设、结果的透明度和全面呈现，以及评估偏倚风险和证据的确定性。概述这些阶段的协议需要进行前瞻性注册。作者应遵循 PRISMA 扩展声明与 NMA 进行系统评价，以确保报告的全面性和透明度。4

一个精心设计的问题对于指导整个 NMA 的作者至关重要，从资格标准的定义到结果的报告，并将有助于确定网络中包括哪些人群和治疗方法，从而形成证据网络可能拿。需要考虑研究问题和基本假设，特别是传递性假设，决定是否应将不同的干预措施评估为单个药物、特定剂量或归为药物类别。5

传递性是指通过中间处理进行间接比较的有效性，是 NMA 的基本假设。它假设纳入的研究在可能影响结果的所有重要因素（例如患者特征和疾病严重程度）方面没有重大差异。例如，涉及共同干预和生物幼稚参与者的试验被排除在系统评价之外，因为它们会导致不及物性。2因此，作者应考虑，例如，与单一疗法相比，已知与更高疗效相关的联合干预试验的资格。

效应修饰符分布的差异在数据中表现为直接估计和间接估计之间的不一致，称为统计不一致，有时也可能是重要异质性的来源。存在几个统计测试，应该用于检查全局（在整个网络中）和本地（在部分网络中）的一致性。如果存在不连贯和/或异质性，则可以使用亚组分析和网络元回归来进一步识别潜在来源。

另一个关键步骤是评估发表偏倚，其中评估小型研究效果是一个重要步骤。这通过称为“比较调整漏斗图”的元分析漏斗图的修改版本进行视觉检查。当图中存在大的不对称性时，小研究效应可能会起作用。网络元回归可以帮助确定原因。此外，应始终计划和进行敏感性分析，以评估结果是否对不同的方法选择具有稳健性，例如排除小型研究或具有高偏倚风险的研究。

清楚地展示研究结果至关重要，而且制作起来可能具有挑战性，尤其是在网络很大的情况下。总体网络效应通常在森林图中报告，而每种处理组合之间的相对效应在排行榜中进行总结。当有许多治疗可用时，二乘二相对效应的数量很快就会变得非常大：例如，在 20 个治疗的网络中，二乘二比较的数量达到了 190个。2 NMA 的一个优点是它的能够提供连贯的治疗排名，其中最受欢迎的指标是累积排名曲线下的表面（SUCRA）。6SUCRA 值介于 0% 和 100% 之间（值越高，治疗排名靠前的可能性越高）。然而，重要的是要结合相对影响结果来解释这些，否则可能会得出误导性结论。例如，英夫利昔单抗、苏金单抗和布罗达单抗的银屑病面积和严重性指数 90 结果的 SUCRA 值分别为 93·6、76·2 和 68·4，2但在比较两者的 2×2 相对效应时另外，由于结果的不确定性很大，这三种药物在疗效上没有显示出显着的统计学差异。因此，排名措施应始终与相对影响一起报告。

此外，已经开发了几种方法来评估从 NMA 获得的证据的确定性。7 , 8 CINeMA（网络元分析置信度：http ://cinema.ispm.ch/ ）是一个扩展 GRADE（推荐评估、开发和评估分级）的网络应用程序，它考虑了六个领域来评估证据的确定性：研究内偏倚、报告偏倚、间接性、不精确性、异质性和不连贯性。使用这些方法对证据的确定性进行评级可以提高结果的透明度、可重复性和可信度。

总之，NMA 复杂且具有挑战性，但如果执行得当，它们可以在比较有效性研究中提供最高水平的证据。在进行 NMA 时，需要在专家临床医生、具有进行系统评价的专业知识的人员以及在 NMA 方面经验丰富的方法学家和统计学家之间进行协作。需要国际努力来鼓励作者和审稿人遵循现有的指导方针，以限制低质量 NMA 的出版。

THE

END