Assessing covariate balance when using the generalized propensity score with quantitative or continuous exposures
Assessing covariate balance when using the generalized propensity score with quantitative or continuous exposures
复制标题
DOI：
10.1177/0962280218756159
复制
发表时间：
2019-05-01
影响因子：
2.3
通讯作者：
Austin, Peter C.
中科院分区：
文献类型：
--
作者：
Austin, Peter C.
在使用观察数据时，倾向评分方法正越来越多地被用于估计治疗和暴露的影响。倾向性评分最初是为二元暴露(例如，积极治疗与对照)而开发的。广义倾向评分是用于定量暴露(例如，药物剂量或数量、收入、受教育年限)的倾向评分的扩展。任何倾向得分分析的一个重要组成部分是平衡评估。这需要评估倾向评分的条件作用(通过匹配、加权或分层)在多大程度上平衡了暴露组之间测量的基线协变量。平衡评估的方法已经得到了很好的描述，并经常在使用具有二元暴露的倾向评分时实施。然而，在使用广义倾向评分时，关于如何评估基线协变量平衡的信息很少。我们描述了当使用广义倾向评分时，基于标准化差异的方法如何适用于定量暴露。我们还描述了一种基于评估定量暴露与样本中每个协变量之间的相关性的方法，当使用基于广义倾向得分的权重进行加权时。我们进行了一系列的蒙特卡罗模拟来评估这些方法的性能。我们还比较了两种不同的估计广义倾向分数的方法：普通最小二乘回归方法和协变量平衡倾向分数方法。我们使用心脏病发作住院患者的数据说明了这些方法的应用，这些患者的定量暴露是肌酐水平。
Propensity score methods are increasingly being used to estimate the effects of treatments and exposures when using observational data. The propensity score was initially developed for use with binary exposures (e.g., active treatment vs. control). The generalized propensity score is an extension of the propensity score for use with quantitative exposures (e.g., dose or quantity of medication, income, years of education). A crucial component of any propensity score analysis is that of balance assessment. This entails assessing the degree to which conditioning on the propensity score (via matching, weighting, or stratification) has balanced measured baseline covariates between exposure groups. Methods for balance assessment have been well described and are frequently implemented when using the propensity score with binary exposures. However, there is a paucity of information on how to assess baseline covariate balance when using the generalized propensity score. We describe how methods based on the standardized difference can be adapted for use with quantitative exposures when using the generalized propensity score. We also describe a method based on assessing the correlation between the quantitative exposure and each covariate in the sample when weighted using generalized propensity score -based weights. We conducted a series of Monte Carlo simulations to evaluate the performance of these methods. We also compared two different methods of estimating the generalized propensity score: ordinary least squared regression and the covariate balancing propensity score method. We illustrate the application of these methods using data on patients hospitalized with a heart attack with the quantitative exposure being creatinine level.