Maclure M
Harvard School of Public Health, Department of Epidemiology, Boston, MA 02115.
Int J Epidemiol. 1990 Dec;19(4):782-7. doi: 10.1093/ije/19.4.782.
Extension of Karl Popper's logic of refutation from the realm of contingency tables to multivariate modelling leads to the conclusion that rigorously scientific multivariate analysis in non-experimental epidemiology differs from the traditional quasi-scientific approach. Instead of aiming for high sensitivity in detecting aetiological agents, the goal in refutation is high specificity--to give the best defence of the 'innocence' of every exposure hypothesized as being a cause. Instead of 'forward selection' or 'backward elimination', multivariate refutation uses the method of 'forward elimination'. This entails a likelihood approach (which may be complemented by, but should be demarcated from, Bayesian methods) not only for statistical inference but also, by analogy, for study design and conduct: one starts with the conclusion (the estimate or hypothesis) and works backwards to the observations (the likelihood of the data or the design of the study). Differences in practice can sometimes be large, as illustrated by a study of hypothesized triggers of myocardial infarction. Multivariate refutation should replace the concept of multivariate modelling in non-experimental epidemiology.
将卡尔·波普尔的证伪逻辑从列联表领域扩展到多变量建模，会得出这样的结论：非实验性流行病学中严格意义上的科学多变量分析不同于传统的准科学方法。证伪的目标不是在检测病因因素时追求高敏感性，而是高特异性——即对每一个被假设为病因的暴露因素的“无辜”进行最佳辩护。多变量证伪不是使用“向前选择”或“向后排除”方法，而是采用“向前排除”法。这需要一种似然方法（可由贝叶斯方法补充，但应与之区分），不仅用于统计推断，而且类推用于研究设计和实施：从结论（估计值或假设）出发，反向推导至观察结果（数据的似然性或研究设计）。实践中的差异有时可能很大，如一项关于心肌梗死假定触发因素的研究所示。在非实验性流行病学中，多变量证伪应取代多变量建模的概念。
