研究人群的代表性:理想和现实之间的取舍

所有的科学研究都是抽样研究,样本代表性是结果外推的前提,对总体或目标人群代表性要求的高低取决于研究问题。在关于一个地区疾病负担等有关问题的研究中,研究者关心的是该地区目前一些具体的事实,总体人群是具体的、明确的、有限的,研究样本的代表性十分重要,抽取有代表性的研究样本也是可行的。相比,在关于病因、疗效、副作用、预后、诊断等问题的研究中,研究者关心的是"放之四海而皆准"的一般规律,总体是模糊的、抽象的、无限的,无法从中抽取有代表性的样本。获得代表性的难度还与观察现象的变异(如交互作用)程度成反比。而且,对于病因和疗效研究,过于强调代表性会导致研究在方法学质量控制方面的妥协,增加偏倚,降低研究结果的内部真实性,使代表性失去意义。因此,这类研究一般不太强调代表性,而是更强调内部真实性,采取的是不断扩展研究人群以重复验证的策略,研究结果最终可代表或可推论的总体就是所有的研究抽样人群的总和所代表的人群。但是,所有研究都必须保证研究人群对于抽样人群的代表性,这是统计推论以及结果真实性和外推性的共同保障。另外,基于大数据的研究也是抽样研究,但是大数据拥有的抽样人群经常是不明确的,因此总体是不明确的、模糊的,推论也必然带着盲目性。

Representativeness of the study population: choice between ideal and reality

All scientific studies are based on samples. The representativeness of the sample is essential for generalization of research findings. The need for high representativeness of the study population depends largely on the nature of the research question. In studies of disease burdens, for example, the researcher's concern is some current specific facts about a specific population, which can be clearly defined and from which drawing a representative sample is both important and possible. In contrast, in studies of causes of disease and effectiveness of treatment, for example, the researcher is interested to find a law of nature in all relevant populations, which are an abstract entity and from which drawing a representative sample is impossible. The difficulty of obtaining representativeness is also inversely related to the variation (like interaction) of the studied phenomenon. Furthermore, overemphasis on representativeness may lead to inevitable compromises in quality control, induce biases, and eventually decrease internal validity, making the gained representativeness compromised. Therefore, research on disease causes and treatment effectiveness relies on repeatedly testing in different populations so as to approach the target population to which the findings can be applied to and which is the totality of the populations represented by the study populations in all the relevant studies. Having said all the above, it is important to note that all studies should draw their samples in a way they represent the population from which the samples are drawn. This forms the basis for the statistical inference and the validity and generalization of epidemiological findings. In addition, any studies based on big data are also sampling studies, in which unfortunately the sampling population is often unclear, which makes generalization of research findings difficult.