Spurious correlations between nuisance features and outcomes present a fundamental challenge for statistical learning and inference. Counterfactual Data Augmentation (CDA), the practice of resampling nuisance features while preserving features of interest, has emerged as a popular strategy to mitigate such correlations, yet its statistical properties, and its reliance on an untestable validity assumption, remain largely unexplored. We provide a statistical treatment of CDA, termed Statistical CDA (SCDA), with two components. First, casting SCDA as a simulated-moment M-estimator, we establish consistency, asymptotic normality with a sandwich covariance and a finite-augmentation-draw variance correction, oracle efficiency for the target parameter, and the equivalence of iterative and one-shot procedures via stochastic approximation, yielding valid large-sample inference where previously none was available. Second, because the augmentation's validity cannot be tested from the observed data when features are entangled (a fact we prove), we develop a sensitivity analysis that bounds the residual nuisance--outcome dependence by a parameter $Γ$, giving sharp identified intervals for the target with valid confidence intervals and a breakdown frontier; the point-identified theory is the $Γ=0$ special case. Simulations confirm the efficiency gains under validity and the robustness bounds under possible invalidity, placing a widely used augmentation method on a transparent inferential footing.

翻译：暂无翻译