We present a new estimator for predicting outcomes in different distributional settings under hidden confounding without relying on instruments or exogenous variables. The population definition of our estimator identifies causal parameters, whose empirical version is plugged into a generative model capable of replicating the conditional law within a test environment. We check that the probabilistic affinity between our proposal and test distributions is invariant across interventions. This work enhances the current statistical comprehension of causality by demonstrating that predictions in a test environment can be made without the need for exogenous variables and without specific assumptions regarding the strength of perturbations or the overlap of distributions.

翻译：我们提出了一种新的估计方法，用于在存在隐藏混杂且不依赖工具变量或外生变量的情况下，预测不同分布设置中的结果。该估计量的总体定义能够识别因果参数，其经验版本被嵌入到一个生成模型中，该模型能够复现测试环境下的条件分布。我们验证了所提方法与测试分布之间的概率亲和性在干预下保持不变。这项工作通过证明可以在无需外生变量、也无需对扰动强度或分布重叠做出特定假设的条件下，在测试环境中进行预测，从而增强了对因果关系的现有统计学理解。