Causal identification typically asks whether a target can be recovered from an infinite observational distribution. In practice, however, analysts only have finite data, and knowing that a quantity is identifiable does not reveal how much data is required for reliable estimation. The paper addresses this gap by introducing Probably Approximately Correct Causal (PACC) learnability, a framework that formalizes uniform accuracy and confidence guarantees over a class of causal models.
The authors derive a general sufficient bound that links sample size to three recurring factors: the strength of causal identification, the complexity of the target estimand, and the complexity of nuisance parameters. This decomposition provides a practical way to reason about finite-sample behavior across different causal inference methods.
To illustrate the framework, the paper analyzes three distinct settings: self-controlled case series, instrumental variables, and propensity-score-based estimation. Each is shown to fit within the same PACC framework, suggesting that diverse causal methods can be compared and studied under a common theoretical lens. The work moves beyond idealized identifiability toward finite-sample guarantees that are more relevant to real-world data analysis.