Guilherme Duarte (Harvard)

Title

Judge-Based IVs: Empirical Tests for Exclusion and Monotonicity, and a Cautious Approach to ATE Bounds and LATE

Abstract

Judge-based instrumental-variable designs are widely used in applied work, yet researchers typically proceed directly to LATE estimates without assessing whether the underlying identification assumptions are compatible with the data. This paper develops a design-first approach that centers empirical testing and partial identification in judge-IV settings. I adapt and extend existing IV inequality tests to assess the exclusion restriction in judge assignments, and I introduce sharper joint tests for monotonicity and exclusion tailored to this design. When the exclusion tests are satisfied, an extension of the Balke–Pearl framework delivers informative partial identification of the average treatment effect (ATE) without imposing monotonicity, and I show that the resulting ATE bounds can be unexpectedly tight and substantially informative in real judge-IV datasets. I then provide targeted sensitivity analyses that quantify the robustness of these conclusions to controlled violations of exclusion. For researchers who also seek complier-specific effects, the paper develops a sharper test of monotonicity—joint with exclusion—than existing approaches, including Frandsen et al. (2023). Passing this test justifies point identification of LATE; failing it triggers a sensitivity framework that characterizes how violations of both monotonicity and exclusion affect complier-specific effects. Together, these tools support a cautious and transparent workflow for judge-based IV designs: empirically test exclusion and monotonicity, obtain informative ATE bounds without unnecessary assumptions, conduct assumption-specific sensitivity analyses, and pursue LATE estimates only when the data support the required identifying conditions.