New Framework Defines "Computational Identifiability" for Causal Effects

Traditional causal identification theory often relies on idealized conditions like infinite data or asymptotic properties to determine if a causal effect can be uniquely computed. This new research proposes a distinct concept called "computational identifiability," which shifts the focus to practical, finite-sample scenarios. Computational identifiability defines a finite computational search procedure to find an empirical estimator for a target query or parameter. If this process successfully identifies an estimator within a specified error tolerance, then the condition is met, contingent on the search assumptions and procedure itself. The framework is demonstrated through experiments, showing its utility in addressing real-world identification challenges, such as those involving small datasets, unclear graphical criteria, mixed data types, and counterfactual estimands.