Ribeiro Andre F
Division of Biosciences, University College London, Gower Street, WC1E, London, United Kingdom.
Department of Applied Mathematics and Statistics, University of Sao Paulo, 13560-970, São Carlos, Brazil.
Sci Rep. 2025 Mar 11;15(1):8423. doi: 10.1038/s41598-024-80839-8.
We study generalization of intervention effects across several simulated and real-world samples. We start by formulating the concept of the 'background' of a sample effect observation. We then formulate conditions for effect generalization based on a sample's set of (observed and unobserved) backgrounds. This reveals two limits for effect generalization: (1) when effects of a variable are observed under all their enumerable backgrounds, or, (2) when backgrounds have become sufficiently randomized. We use the resulting combinatorial framework to re-examine open issues in current causal effect estimators: out-of-sample validity, concurrent estimation of multiple effects, bias-variance tradeoffs, statistical power, and connections to current predictive and explaining techniques. Methodologically, these definitions also allow us to replace the parametric estimation problems that followed the 'counterfactual' definition of causal effects by combinatorial enumeration and randomization problems in non-experimental samples. We use the resulting non-parametric framework to demonstrate (External Validity, Unconfoundness and Precision) tradeoffs in the performance of popular supervised, explaining, and causal-effect estimators.
我们研究了干预效应在多个模拟样本和真实世界样本中的泛化情况。我们首先阐述了样本效应观察“背景”的概念。然后，我们基于样本的（观察到的和未观察到的）背景集制定了效应泛化的条件。这揭示了效应泛化的两个限制：（1）当一个变量的效应在其所有可枚举的背景下都被观察到时；或者，（2）当背景已经充分随机化时。我们使用由此产生的组合框架来重新审视当前因果效应估计器中的开放性问题：样本外有效性、多种效应的并发估计、偏差 - 方差权衡、统计功效以及与当前预测和解释技术的联系。从方法上讲，这些定义还使我们能够通过组合枚举和非实验样本中的随机化问题，取代遵循因果效应“反事实”定义的参数估计问题。我们使用由此产生的非参数框架来展示流行的监督、解释和因果效应估计器在性能方面的（外部有效性、无混杂性和精度）权衡。
