Department of Medical Biometry and Bioinformatics, Heinrich Heine University Düsseldorf, Düsseldorf, Germany.
Institute of Rheumatology, Medical Faculty, Heinrich Heine University Düsseldorf, Düsseldorf, Germany.
Stat Med. 2024 Jul 30;43(17):3140-3163. doi: 10.1002/sim.10110. Epub 2024 May 27.
Weighting methods are widely used for causal effect estimation in non-randomised studies. In general, these methods use the propensity score (PS), the probability of receiving the treatment given the covariates, to arrive at the respective weights. All of these "modelling" methods actually optimize prediction of the respective outcome, which is, in the PS model, treatment assignment. However, this does not match with the actual aim of weighting, which is eliminating the association between covariates and treatment assignment. In the "balancing" approach, covariates are thus balanced directly by solving systems of numerical equations, explicitly without fitting a PS model. To compare modelling, balancing and hybrid approaches to weighting we performed a large simulation study for a binary treatment and a survival outcome. For maximal practical relevance all simulation parameters were selected after a systematic review of medical studies that used PS methods for analysis. We also introduce a new hybrid method that uses the idea of the covariate balancing propensity score and matching weights, thus avoiding extreme weights. In addition, we present a corrected robust variance estimator for some of the methods. Overall, our simulations results indicate that balancing approach methods work worse than expected. However, among the considered balancing methods, entropy balancing consistently outperforms the variance balancing approach. All methods estimating the average treatment effect in the overlap population perform well with very little bias and small standard errors even in settings with misspecified propensity score models. Finally, the coverage using the standard robust variance estimator was too high for all methods, with the proposed corrected robust variance estimator improving coverage in a variety of settings.
权重法广泛应用于非随机研究中的因果效应估计。一般来说，这些方法使用倾向得分（PS），即给定协变量时接受治疗的概率，来得到相应的权重。所有这些“建模”方法实际上都是优化对各自结果的预测，在 PS 模型中，这是治疗分配。然而，这与加权的实际目的不符，加权的目的是消除协变量与治疗分配之间的关联。在“平衡”方法中，通过求解数值方程组直接平衡协变量，明确地不拟合 PS 模型。为了比较建模、平衡和混合加权方法，我们针对二分类治疗和生存结局进行了大规模模拟研究。为了最大限度地提高实际相关性，所有模拟参数都是在对使用 PS 方法进行分析的医学研究进行系统回顾后选择的。我们还引入了一种新的混合方法，该方法使用协变量平衡倾向得分和匹配权重的思想，从而避免极端权重。此外，我们还为一些方法提出了一种修正后的稳健方差估计量。总体而言，我们的模拟结果表明，平衡方法的效果不如预期。然而，在考虑的平衡方法中，熵平衡始终优于方差平衡方法。在重叠人群中估计平均治疗效果的所有方法都表现良好，即使在 PS 模型指定不正确的情况下，也只有很小的偏差和小的标准误差。最后，所有方法使用标准稳健方差估计量的覆盖范围都过高，而提出的修正后的稳健方差估计量在多种情况下都提高了覆盖范围。
