Department of Epidemiology and Biostatistics, School of Public Health, Tehran University of Medical Sciences, Tehran, Iran.
Knowledge Utilization Research Center, Tehran University of Medical Sciences, Tehran, Iran.
Arch Iran Med. 2018 Apr 1;21(4):164-169.
The goal of many observational studies is to estimate the causal effect of an exposure on an outcome after adjustment for confounders, but there are still some serious errors in adjusting confounders in clinical journals. Standard regression modeling (e.g., ordinary logistic regression) fails to estimate the average effect of exposure in total population in the presence of interaction between exposure and covariates, and also cannot adjust for time-varying confounding appropriately. Moreover, stepwise algorithms of the selection of confounders based on P values may miss important confounders and lead to bias in effect estimates. Causal methods overcome these limitations. We illustrate three causal methods including inverse-probability-of-treatment-weighting (IPTW) and parametric g-formula, with an emphasis on a clever combination of these 2 methods: targeted maximum likelihood estimation (TMLE) which enjoys a double-robust property against bias.
许多观察性研究的目的是在调整混杂因素后估计暴露对结局的因果效应，但在临床期刊中调整混杂因素仍存在一些严重错误。标准回归建模（例如普通逻辑回归）在暴露与协变量之间存在交互作用的情况下无法估计总人群中暴露的平均效应，也无法适当地调整时变混杂因素。此外，基于 P 值的混杂因素选择的逐步算法可能会遗漏重要的混杂因素，并导致效应估计的偏差。因果方法克服了这些局限性。我们展示了三种因果方法，包括治疗逆概率加权（IPTW）和参数 g 公式，重点介绍了这两种方法的巧妙组合：具有双重稳健性的靶向最大似然估计（TMLE）。
