Resampling-based inference for causal effect estimates in time-to-event data

基于重采样的事件时间数据因果效应估计推断

• 批准号: 439942859
• 负责人: Professor Dr. Jan Beyersmann
• 金额: --
• 依托单位: Institut für Mathematik
• 依托单位国家: 德国
• 项目类别: Research Grants
• 资助国家: 德国
• 项目状态: 未结题
• 来源: https://gepris.dfg.de/gepris/projekt/439942859?language=en
• 关键词: Resampling; based; inference; causal; effect

Drawing causal conclusions from data becomes more and more relevant, not only in the context of growing data bases and therefore increasing use of observational data, but also in randomized controlled trials, where analyses often need to be adjusted for (despite randomization) unbalanced covariates or time-varying exposures arising from, e.g., a lack of adherence. The situation is even more complex, when the outcome is time-to-event, possibly subject to competing risks. During the first phase of the project, we rigorously investigated different resampling schemes for the estimation of confidence intervals and confidence bands of the average treatment effect in a setting with competing risks and right-censored data. It is the aim of this renewal proposal to extend the results in several directions: First, the issue of left-truncation has so far received little attention in causal inference literature. It is, however, relevant in several scenarios, reaching from pregnancy studies to time scales such as age or calendar time. Second, a common approach in causal inference is matching, e.g. based on propensity scores. Since this results in dependent data, the nonparametric bootstrap is no longer valid, but wild bootstrap approaches might still work, as has been shown in settings with a continuous or binary outcome, but not in the context of survival data, before. Third, a more complicated situation arises, when time-dependent confounding is present. We also aim to develop methods for this context, investigating landmark approaches and extending previous work on exposure density sampling and nested exposure case-control sampling. All methods will be available to a broad scientific audience by providing R code. This also increases both transparency and replicability of published simulation results. Moreover, simulated data shall be supplemented with empirical studies based on real data to increase the explanatory power of the results.

从数据中得出因果结论变得越来越重要，不仅在数据库不断扩大的背景下，因此也越来越多地使用观测数据，而且在随机对照试验中也是如此，在随机对照试验中，分析经常需要针对(尽管随机化)协变量不平衡或因缺乏遵循而产生的时变暴露进行调整。情况甚至更加复杂，当结果是事件发生的时间，可能会受到相互竞争的风险的影响。在项目的第一阶段，我们严格研究了不同的重抽样方案，以估计在竞争风险和右删失数据的情况下平均治疗效果的可信区间和可信区间。这个更新建议的目的是在几个方向上扩展结果：第一，到目前为止，在因果推理文献中，左截断问题几乎没有受到关注。然而，它在几个场景中都是相关的，从怀孕研究到年龄或日历时间等时间尺度。第二，因果推理中的一种常见方法是匹配，例如基于倾向分数。由于这会导致相依数据，因此非参数自举方法不再有效，但野生自举方法可能仍然有效，如之前在具有连续或二元结果的设置中所示，但不是在生存数据的情况下。第三，当存在依赖时间的混淆时，会出现更复杂的情况。我们还致力于为这一背景开发方法，调查具有里程碑意义的方法，并扩展先前在暴露密度抽样和嵌套暴露病例对照抽样方面的工作。通过提供R代码，所有方法都将提供给广泛的科学受众。这还增加了已发布的模拟结果的透明度和可复制性。此外，模拟数据应辅以以真实数据为基础的实证研究，以增加结果的解释力。