Departments of Ophthalmology and Visual Sciences, Medicine and Pharmacology, University of British Columbia, Vancouver, British Columbia.
Department of Epidemiology and Medicine, McGill University, Montreal, Canada.
Am Heart J. 2021 Jul;237:62-67. doi: 10.1016/j.ahj.2021.03.008. Epub 2021 Mar 17.
Covariate adjustment is integral to the validity of observational studies assessing causal effects. It is common practice to adjust for as many variables as possible in observational studies in the hopes of reducing confounding by other variables. However, indiscriminate adjustment for variables using standard regression models may actually lead to biased estimates. In this paper, we differentiate between confounders, mediators, colliders, and effect modifiers. We will discuss that while confounders should be adjusted for in the analysis, one should be wary of adjusting for colliders. Mediators should not be adjusted for when examining the total effect of an exposure on an outcome. Automated statistical programs should not be used to decide which variables to include in causal models. Using a case scenario in cardiology, we will demonstrate how to identify confounders, colliders, mediators and effect modifiers and the implications of adjustment or non-adjustment for each of them.
协变量调整是评估因果效应的观察性研究有效性的组成部分。在观察性研究中，常见的做法是尽可能调整更多的变量，希望减少其他变量引起的混杂。然而，使用标准回归模型不加区分地调整变量实际上可能导致有偏估计。在本文中，我们区分了混杂因素、中介因素、共发因素和效应修饰因素。我们将讨论到，虽然在分析中应该调整混杂因素，但应该警惕调整共发因素。在检查暴露对结果的总效应时，不应调整中介因素。自动化统计程序不应用于决定因果模型中应包含哪些变量。我们将使用心脏病学中的一个案例场景来演示如何识别混杂因素、共发因素、中介因素和效应修饰因素，以及对它们进行调整或不调整的影响。
