再论观察与实验：大数据现实世界研究不能取代随机对照试验

传统上，流行病学多以干预划分观察和实验，干预研究等于实验研究，还认为干预研究的科学性高于观察性研究。在一般科学实验里，干预指人为施加的改变自然状况的措施。干预并不一定是有益的，也并不一定是研究者当下施加的，研究者、受试者或第三者目前或过去施加的措施都可以形成"有效的"干预。例如，由研究者、受试者和第三者通过某种方法致使视神经损伤，都可以形成有效改变视神经正常功能的干预，研究者可以由此观察到视神经和视力的关系。以此推论，由受试者自己过去施加的不良干预（如吸烟）也属于干预，那么研究吸烟和肺癌的观察性队列研究就等同于实验研究了。由此看来，干预本身并不足以有效地区分观察和实验。如果认为实验的科学性高于观察，那么在干预的基础上，只能从科学性上（即设计特征）区分观察和实验。在评估医学干预效果的临床试验中，随机分组是在传统认为的观察研究基础上引入的最重要的偏倚控制措施，应该是区分观察和实验的核心属性。如果一定要把人群研究分成观察和实验，随机对照试验才是真正的实验研究，非随机分组形成的干预研究属于试验，但不是实验。基于大数据的现实世界研究，如果没有随机分组，不能构成实验，也不能成为对干预效果的最终检验。大数据现实世界研究不能取代随机对照试验，这是本文希望传达的最重要的信息。

Revisit to observation and experiment: real world study cannot replace randomized controlled trial

In epidemiology, intervention is normally used to define what experiment is intervention studies are equaled to experimental studies. Experimental studies are also considered scientifically more rigorous than observational ones. Intervention is generally referred to human activities that can interfere or change natural conditions. The intervention by definition may not necessarily be beneficial to the study subjects (although exposing harmful interventions to humans are unethical) and activities by the researcher, by the subject himself, or by any third party and either now or in the past can all form "effective" interventions. For example, interventions that can damage the optic nerve by any of the three parties can all help the researcher establish the relation between the optic nerve and vision. In the same sense, an activity that a study subject initiated in the past, such as smoking, would also constitute a valid intervention. As a result, a cohort study on smoking and lung cancer would also be an experiment. From the above arguments, we can see that intervention alone does not suffice to distinguish between experiment and observation. As we equal experiment to higher scientific rigorousness than observation, only can study designing features of intervention studies be used to define experiment. In intervention trials, randomization is the defining feature that makes randomized controlled trials differ from, and scientifically more rigorous than, controlled observational studies and has been commonly used to define experiment. If we have to divide clinical research into experiment and observation, randomized controlled trials would be experimental and non-randomized studies of intervention are trials but not experiment. Big data, real-world studies are not experiment and cannot replace randomized trials in confirmation of efficacy if comparison groups are not formed by randomization. Real world studies cannot replace randomized controlled trials. This is the most important message this paper wishes to convey.