NYU Researchers Propose 'Computational Identifiability' — and It Could Change How We Do Causal Inference

Traditional causal identifiability relies on theoretical notions that assume infinite data and asymptotic properties, making it impractical for finite-sample real-world settings. In their new paper, researchers Lucius Bynum, Rajesh Ranganath, and Kyunghyun Cho propose 'computational identifiability,' which replaces these idealized conditions with a finite computational search procedure. The process searches for an empirical estimator that meets a specified error tolerance, conditional on a prior distribution over parameters and the search procedure itself. This approach allows researchers to answer practical identification questions that classical theory cannot handle, such as identification with very small samples, ambiguous causal graphs, or mixed data types (observational plus interventional). The framework also extends to counterfactual queries.

The authors demonstrate through experiments that computational identifiability works in scenarios where theoretical identifiability fails. For example, they show it can identify effects from datasets with only dozens of samples, with partially specified causal graphs, or when combining experimental and observational data. This bridges the gap between theory and application, making causal inference more accessible for fields like epidemiology, economics, and AI where data is limited or messy. The paper includes open-source code (available at the provided GitHub link) to implement the search procedure. By grounding identifiability in computation rather than mathematical idealization, the work offers a practical toolkit for causal discovery under real constraints.