They formalized an econometric framework for analyzing natural and imperfectly randomized experiments. The core insight interprets two-stage least squares through the lens of the Local Average Treatment Effect (LATE). Researchers can now estimate causal impacts even when participation is voluntary. Applications span education, health, and public policy, greatly increasing transparency. Their methods are now standard in textbooks on causal inference.

Angrist and Imbens (1994) reinterpreted 2SLS estimators in experiments with probabilistic compliance. Using the potential-outcomes framework, they showed that LATE equals the ratio of reduced-form to first-stage coefficients, LATE = β_RF / β_FS. Crucially, the estimate pertains only to “compliers”—individuals whose behavior shifts because of the instrument—so external validity must be assessed separately. They formalized the assumptions of random assignment, exclusion, and monotonicity, urging researchers to test them explicitly. The framework revolutionized instrumental-variables reasoning and now informs field experiments, Mendelian randomization, and more.

Imbens & Angrist (1994) supplied an identification theorem for LATE given IID instruments, exclusion, and monotonicity, linking Wald ratios to complier ATEs. Follow-up work (Angrist, Imbens & Rubin 1996) adopted a Bayesian randomization lens, facilitating finite-sample inference and sensitivity checks. Their insights precipitated extensions like marginal treatment effects, weak-instrument diagnostics, and double-machine-learning estimators for high-dimensional covariates. Today the IV-LATE framework underpins policy simulation, extrapolation kernels, and synthetic-control methodologies.