Drake C, Fisher L
Division of Statistics, University of California, Davis 95616, USA.
Int J Epidemiol. 1995 Feb;24(1):183-7. doi: 10.1093/ije/24.1.183.
Subjects in observational studies of exposure effects have not been randomized to exposure groups and may therefore differ systematically with regard to variables related to exposure and/or outcome. To obtain unbiased estimates and tests of exposure effects one needs to adjust for these variables. A common method is adjustment via a parametric model incorporating all known prognostic variables. Rosenbaum and Rubin propose adjustment by the conditional exposure probability given a set of covariates which they call the propensity score. They show that, at any value of the propensity score, covariates are on average balanced between exposure groups. Thus matching on the propensity score leads to unbiased estimators and tests of exposure effect. However, the validity of the method depends on knowing the exposure probability. This quantity is usually not known in observational studies and needs to be estimated.
暴露效应观察性研究中的受试者并未被随机分配到暴露组，因此在与暴露和/或结果相关的变量方面可能存在系统性差异。为了获得无偏的暴露效应估计值和检验结果，需要对这些变量进行调整。一种常用的方法是通过纳入所有已知预后变量的参数模型进行调整。罗森鲍姆和鲁宾提出根据一组协变量的条件暴露概率进行调整，他们将其称为倾向得分。他们表明，在倾向得分的任何值上，协变量在暴露组之间平均是平衡的。因此，根据倾向得分进行匹配可得到无偏的暴露效应估计值和检验结果。然而，该方法的有效性取决于是否知道暴露概率。在观察性研究中，这个量通常是未知的，需要进行估计。
