rct

This paper presents a partial-identification approach to use prior RCT data to bound the causal effect of a new ML model, leveraging assumptions about counterfactual correctness and subgroup predictive accuracy to yield more informative bounds.