A comparison of entropy balance and probability weighting methods to generalize observational cohorts to a population: a simulation and empirical example
A comparison of entropy balance and probability weighting methods to generalize observational cohorts to a population: a simulation and empirical example
复制标题
DOI：
10.1002/pds.4121
复制
发表时间：
2017-04-01
影响因子：
2.6
通讯作者：
Huang, Joanna C.
中科院分区：
文献类型：
--
作者：
Harvey, Raymond A.;Hayden, Jennifer D.;Huang, Joanna C.
目的比较观察性研究中控制偏倚和混杂的方法，包括逆概率加权法（IPW）和稳定化IPW法（sIPW）。这些方法通常需要迭代和后校准来实现协变量平衡。相比之下，熵平衡（EB）优化协变量平衡先验校准权重使用目标的时刻作为constrains.MethodsWe测量协变量平衡经验和模拟使用绝对标准化均值差（ASMD），绝对偏差（AB），和均方根误差（RMSE），调查两种情况：观察（暴露）队列的大小超过目标（未暴露）队列，反之亦然。实证应用加权的商业健康计划队列的全国代表性的国家健康和营养检查调查目标相同的协变量和比较平均总医疗保健成本估计跨methods.ResultsEntropy平衡单独实现平衡（ASMD 0.10）在模拟和经验的所有协变量。在模拟情景I中，EB实现了最低的AB和RMSE（13.64，31.19），与IPW（263.05，263.99）和sIPW（319.91，320.71）相比。在情景II中，EB表现优于IPW和sIPW，AB和RMSE较小。在情景I和II中，EB分别与IPW和sIPW相比，与模拟人群结局（490.05美元，487.62美元）的平均估计差异最低。经验上，只有EB不同于未加权的平均成本表明IPW，和sIPW weighting is effective.ConclusionEntropy balance演示的偏差方差权衡实现更高的估计精度，但估计精度较低，相比IPW方法。EB加权不需要后处理，并有效减轻了观察到的偏倚和混淆。版权所有（c）2016约翰威利父子有限公司
PurposeWe compared methods to control bias and confounding in observational studies including inverse probability weighting (IPW) and stabilized IPW (sIPW). These methods often require iteration and post-calibration to achieve covariate balance. In comparison, entropy balance (EB) optimizes covariate balance a priori by calibrating weights using the target's moments as constraints.MethodsWe measured covariate balance empirically and by simulation by using absolute standardized mean difference (ASMD), absolute bias (AB), and root mean square error (RMSE), investigating two scenarios: the size of the observed (exposed) cohort exceeds the target (unexposed) cohort and vice versa. The empirical application weighted a commercial health plan cohort to a nationally representative National Health and Nutrition Examination Survey target on the same covariates and compared average total health care cost estimates across methods.ResultsEntropy balance alone achieved balance (ASMD0.10) on all covariates in simulation and empirically. In simulation scenario I, EB achieved the lowest AB and RMSE (13.64, 31.19) compared with IPW (263.05, 263.99) and sIPW (319.91, 320.71). In scenario II, EB outperformed IPW and sIPW with smaller AB and RMSE. In scenarios I and II, EB achieved the lowest mean estimate difference from the simulated population outcome ($490.05, $487.62) compared with IPW and sIPW, respectively. Empirically, only EB differed from the unweighted mean cost indicating IPW, and sIPW weighting was ineffective.ConclusionEntropy balance demonstrated the bias-variance tradeoff achieving higher estimate accuracy, yet lower estimate precision, compared with IPW methods. EB weighting required no post-processing and effectively mitigated observed bias and confounding. Copyright (c) 2016 John Wiley & Sons, Ltd.