Evaluating Counterfactual Policies with Instruments

This paper addresses the problem of evaluating the impact of counterfactual policies, like changing treatment assignment, using instrumental variables. It provides a computationally efficient framework for bounding the effects of such policies, without relying on the often-restrictive monotonicity assumption. The work is significant because it offers a more robust approach to policy evaluation, especially in scenarios where traditional IV methods might be unreliable. The applications to real-world datasets (bail judges and prosecutors) further enhance the paper's practical relevance.