临床流行病学的关键概念：解决和报告随机对照试验中的偏倚来源

1. Background

Randomized controlled trials are widely considered the most robust design for evaluating the effects of clinical interventions because randomization can potentially eliminate bias resulting from differences in pre-existing characteristics of participants – prognostic factors in particular – in intervention and comparator conditions. If designed and conducted well, trials have high internal validity, meaning that inferences of causal relationships (ie, that an intervention causes a change in outcome) are free of systematic error (or bias) [1]. However, in many areas of clinical research practical issues with trials can compromise the integrity of randomization and lead to bias. In addition, there remain sources of bias that cannot be addressed by randomization and which can occur during the whole process of research (during design, conduct, analysis, and report of trials) [2–4]. Such bias can reduce the internal validity of a trial, leading to a distortion of the true treatment effect [3]. Importantly, the risk of bias may differ for different outcomes within the same trial.

The revised Cochrane risk-of-bias tool for trials (RoB 2) distinguishes five domains of bias that can affect the results of trials stemming from (1) the randomization process, (2) deviations from intended interventions, (3) missing outcome data; (4) outcome measurement, and (5) reporting of findings (www.riskofbias.info) [3,5]. The tool was designed to assess the risk of bias of trials included in systematic reviews (other tools exist [6]). Here, we use RoB 2 as a framework for recommendations to help researchers planning or conducting trials to mitigate these sources of bias and ensure transparency in reporting so that users of research are aware of them. It may also provide readers of trial reports with guidance for critical appraisal of trial reports.

2. Sources of bias and recommendations for prevention and mitigation

Bias can arise from the randomization process when unreliable methods are used for generating the random allocation sequence, when treatment allocation is not adequately concealed, or when the randomization process is not well implemented. This can result in potentially confounding variables being unbalanced between trial arms at baseline [7]. Lack of concealment of treatment assignment (blinding) often cannot be avoided and does not necessarily lead to bias but it should be reported so that risk of bias can be adequately assessed.

Bias due to deviations from intended interventions can arise when treatments are not delivered fully as intended (lack of fidelity). It can also arise from exposure of participants to factors influencing the outcome other than the intervention to which they have been assigned, including accidental exposure to an intervention in another trial arm. This “contamination” can arise, for example, when participants fail to receive a hoped-for benefit from their intervention and seek an alternative treatment. Bias can also arise from participants not adhering fully to the treatment regimen. This is very common and often cannot be avoided, but level of adherence should always be assessed and reported. Finally, bias can arise because those delivering an intervention in one arm are more enthusiastic about their intervention than those in another arm. This “allegiance bias” can affect fidelity and even within the parameters set for delivery of the intervention, can inflate or reduce the effect size [8].

Missing outcome data are an ubiquitous problem in clinical trials, often because participants no longer wish to engage with the trial. If outcome data are missing “at random,” that is, in a way that is unrelated to the outcome, the effect is only to reduce the statistical power because of reduced sample size for analysis. However, if outcome data are missing in a way that could be related to the outcome of interest, this breaks the randomization and undermines the key strength of the trial design [9]. If outcome data are missing to different extents, or for different reasons, in different arms of the trial this can lead to an underestimate or overestimate of any effects.

Bias in outcome measurement can be due many factors, such as expectations favoring an intervention or incentives to produce data that confirm predictions. Clinical trials often rely on subjective outcome measures that are potentially subject to bias and error. Differential bias can arise from different methods being used to assess outcomes in different trials arms. Bias can also arise from repeated testing or changes in reference points used to judge outcomes (“response shift”).

Bias in selection of findings to report or highlight is common and not fully addressed by requirements to register study protocols or analysis plans prior to analyzing the data. Subgroup analyses are particularly vulnerable to this bias as they provide multiple opportunities to find and select or highlight findings that accord with a desired outcome [10]. Underpowered studies are also particularly vulnerable to reporting bias, compounded by a tendency of journals to be more inclined to accept articles reporting positive findings. The funding source or sponsor of a study can also be a factor contributing to reporting bias.

Table 1 provides an overview of the sources of bias outlined, together with recommendations for prevention, mitigation, and reporting. Regarding the latter, a trial report should be written in accordance with the CONSORT (Consolidated Standards of Reporting Trials) guidelines and relevant extensions (see for a full list of reporting guidelines: www.equator-network.org) [11]. A new, free online Paper Authoring Tool (https://paperauthoringtool.com) has used these guidelines and extensive experience in writing up clinical trials to provide authors with a way to ensure that trial protocols and reports are prepared in a way that maximises transparency. We recommend to report any additional measures that were taken to reduce the risk of bias, for example, how treatment preferences, adherence and fidelity were enhanced. The trial report should provide data that have been collected in relation to the different sources of bias, if any, and results from analyses incorporating these (eg, results from sensitivity analyses). Finally, authors should not only discuss if a relevant form of bias could have been introduced but also how this would affect the interpretation of the study findings.

3. Conclusion

While randomized controlled trials can provide a high degree of confidence that outcomes are caused by interventions, in practice there are major, often unavoidable, sources of bias. Trialists must pay close attention to all of these and take steps to mitigate them where possible and always to report them to allow users of research to form a judgement about the extent to which findings can be relied upon. Our concise overview of important sources of bias – based on Cochrane's risk-of-bias (RoB 2) framework – together with our list of recommendations for their prevention, mitigation, and reporting may provide guidance in this respect.

全文翻译（仅供参考）

随机对照试验被广泛认为是评估临床干预效果的最稳健设计。虽然普遍性通常是有限的，但随机化旨在确保观察到的效果是真实的。然而，即使在进行良好的试验中，也存在常见的偏见来源，这对这种解释构成了威胁。修订后的 Cochrane 试验偏倚风险工具 (RoB 2) 区分了五个可能影响试验结果的偏倚领域，这些偏倚源自 (1) 随机化过程，(2) 与预期干预措施的偏差，(3) 缺少结果数据，（4）结果测量，和（5）结果报告。我们使用 RoB 2 作为建议框架，以帮助研究人员减轻这些偏见来源并确保报告的透明度，以便研究用户了解它们。

1 . 背景

随机对照试验被广泛认为是评估临床干预效果的最稳健设计，因为随机化可以潜在地消除由于参与者预先存在的特征（尤其是预后因素）在干预和比较条件下的差异而导致的偏倚。如果设计和实施得当，试验具有很高的内部有效性，这意味着因果关系的推论（即干预导致结果的变化）没有系统错误（或偏差）[1]. 然而，在临床研究的许多领域中，试验的实际问题可能会损害随机化的完整性并导致偏倚。此外，仍然存在无法通过随机化解决的偏倚来源，并且可能在整个研究过程中发生（在设计、实施、分析和试验报告期间）[ 2-4 ]。这种偏差会降低试验的内部有效性，导致真实治疗效果的失真[3]。重要的是，同一试验中不同结果的偏倚风险可能不同。

修订后的 Cochrane 试验偏倚风险工具 (RoB 2) 区分了五个可能影响试验结果的偏倚领域，这些偏倚源自 (1) 随机化过程，(2) 与预期干预措施的偏差，(3) 缺少结果数据; (4) 结果测量，和 (5) 报告结果[ 3 , 5 ]。该工具旨在评估系统评价中包含的试验的偏倚风险（存在其他工具[6]）。在这里，我们使用 RoB 2 作为建议框架，以帮助研究人员计划或进行试验，以减轻这些偏见来源并确保报告的透明度，以便研究用户了解它们。它还可以为试验报告的读者提供对试验报告进行批判性评价的指导。

2 . 偏见的来源以及预防和缓解的建议

当使用不可靠的方法来生成随机分配序列时，当治疗分配没有充分隐藏时，或者当随机化过程没有很好地实施时，随机化过程可能会产生偏差。这可能导致基线时试验组之间的潜在混杂变量不平衡[7]。治疗分配（盲法）的缺乏隐藏通常无法避免，也不一定会导致偏倚，但应报告，以便充分评估偏倚风险。

当治疗未完全按预期进行时（缺乏保真度），可能会因偏离预期干预而产生偏差。它也可能源于参与者暴露于影响结果的因素而不是他们被分配的干预措施，包括意外暴露于另一个试验组中的干预措施。例如，当参与者未能从他们的干预中获得预期的好处并寻求替代治疗时，就会出现这种“污染”。偏差也可能源于参与者没有完全遵守治疗方案。这是非常常见的，通常是无法避免的，但应始终评估和报告依从性水平。最后，可能会出现偏见，因为那些在一个手臂上进行干预的人比在另一个手臂上的人更热衷于他们的干预。[8] .

缺少结果数据是临床试验中普遍存在的问题，通常是因为参与者不再希望参与试验。如果结果数据“随机”缺失，即以与结果无关的方式缺失，其效果只是降低了统计功效，因为分析的样本量减少了。但是，如果结果数据以可能与感兴趣的结果相关的方式丢失，这会破坏随机化并破坏试验设计的关键强度[9]。如果结果数据在不同程度上或由于不同原因在试验的不同组中缺失，这可能导致低估或高估任何影响。

结果测量中的偏差可能是由于许多因素造成的，例如有利于干预的期望或产生证实预测的数据的动机。临床试验通常依赖于可能存在偏差和错误的主观结果测量。差异偏倚可能源于用于评估不同试验组结果的不同方法。偏差也可能来自重复测试或用于判断结果的参考点的变化（“响应转移”）。

在选择报告或强调的结果时存在偏差是常见的，并且在分析数据之前注册研究方案或分析计划的要求并未完全解决。亚组分析特别容易受到这种偏见的影响，因为它们提供了多种机会来发现和选择或突出符合预期结果的结果[10]。动力不足的研究也特别容易受到报告偏见的影响，而期刊更倾向于接受报告积极发现的文章。研究的资金来源或赞助商也可能是导致报告偏倚的一个因素。

关于后者，应根据 CONSORT（报告试验综合标准）指南和相关扩展编写试验报告[11]。一个新的、免费的在线论文创作工具在撰写临床试验时使用了这些指南和丰富的经验，为作者提供了一种方法，以确保以最大限度地提高透明度的方式准备试验方案和报告。我们建议报告为降低偏倚风险而采取的任何其他措施，例如，如何提高治疗偏好、依从性和保真度。试验报告应提供已收集到的与不同偏倚来源（如果有）相关的数据，以及包含这些偏倚的分析结果（例如，敏感性分析的结果）。最后，作者不仅应该讨论是否可以引入相关形式的偏见，还应该讨论这将如何影响对研究结果的解释。

3 . 结论

虽然随机对照试验可以提供对结果由干预措施引起的高度置信度，但在实践中存在主要的、通常是不可避免的偏倚来源。试验者必须密切关注所有这些问题，并在可能的情况下采取措施减轻这些问题，并始终报告这些问题，以便研究用户对研究结果的可信赖程度形成判断。基于 Cochrane 的偏倚风险 (RoB 2) 框架，我们对重要偏倚来源的简明概述，以及我们的预防、缓解和报告建议列表，可能会在这方面提供指导。