IV

Instrumental Variables MOC

Instrumental variables (IV) identify causal effects by isolating variation in a treatment that is driven by an external instrument — variation that is plausibly as-good-as-randomly assigned and affects the outcome only through the treatment. The framework’s defining insight is interpretive: with heterogeneous effects, IV does not recover an overall average but the LATE — the effect for compliers, the units the instrument actually moves. The identifying triad is instrument independence (Randomization), the Exclusion-Restriction, and Monotonicity, plus a first stage.

Papers

• Angrist2022-EmpiricalStrategiesIlluminatingPath — Angrist’s Nobel lecture: the LATE framework, separating cheap independence from the substantive exclusion restriction, illustrated with school-admission instruments.
• AngristKrueger1999-EmpiricalStrategiesLaborEconomics — the handbook chapter that systematized IV (and other empirical strategies) in labor economics.
• AngristKrueger2001-SearchForIdentification — JEP history of IV from supply–demand systems to natural experiments; the exclusion restriction as the binding assumption.
• Imbens2014-IVEconometriciansPerspective — potential-outcomes synthesis bridging the econometric and statistical accounts; IV as a complier-specific LATE.
• Murray2006-WeakAndInvalidInstruments — practitioner’s guide to the two central threats: invalid instruments (exclusion) and weak instruments (first stage).
• RosenzweigWolpin2000-NaturalNaturalExperiments — structuralist critique: natural experiments still need an economic model to interpret what is identified.
• DeChaisemartinDHaultfoeuille2018-FuzzyDiD — fuzzy DiD as an IV/LATE problem in a panel setting; bridges DiD and IV.
• FreyaldenhovenEtAl2019-PreEventTrendsPanelEventStudy — IV inside the event study: policy leads as excluded instruments, an unaffected covariate as the endogenous proxy for the confound; bridges DiD’s pre-trends problem and IV.
• AngristPischke2010-CredibilityRevolution — situates IV within the broader design-based credibility revolution.

Shift-share / Bartik instruments

• BlanchardKatz1992-RegionalEvolutions — origin (with Bartik 1991) of the industry-mix Shift-Share-Instrument: predicted state employment growth from national industry growth × local shares, used to identify labor-demand shocks.
• Card2001-ImmigrantInflows — the ethnic-enclave supply-push instrument: predicted immigrant inflows from historical origin-country settlement shares × national inflows.
• AutorDornHanson2013-ChinaSyndrome — the leading modern application: local import exposure instrumented by Chinese exports to other high-income countries × initial industry shares.
• GoldsmithPinkhamEtAl2020-BartikInstruments — Bartik IV as a share-exogeneity design; Rotemberg weights decompose the estimate and diagnose it.
• BorusyakEtAl2022-QuasiExperimentalShiftShare — the counterpart shock-exogeneity design: quasi-random shifters with (possibly) endogenous shares.
• AdaoEtAl2019-ShiftShareDesigns — inference: shared shares correlate residuals across regions, so conventional standard errors over-reject; derives valid (AKM) ones.

Key Concepts

LATE · Exclusion-Restriction · Monotonicity · Randomization (instrument independence) · Weak-Instruments (relevance / first stage) · Shift-Share-Instrument (Bartik) · Design-Based-Inference · Treatment-Effect-Heterogeneity · Causal-Estimand · SUTVA

Debates & Contradictions

• What does IV estimate? The LATE revolution (Imbens–Angrist, here via Angrist 2022) reframed IV as recovering a complier-specific effect, not a universal ATE — a shift from the older “IV estimates the structural parameter” view that some structural econometricians still contest.
• Independence vs exclusion. Angrist stresses that random assignment buys independence cheaply, but the exclusion restriction is a separate, untestable, substantive commitment — the assumption most IV critiques actually target.
• External validity. A complier-defined estimand is internally credible but may not generalize to always-takers or never-takers — the recurring price of design- based identification.
• Design vs. structure. RosenzweigWolpin2000-NaturalNaturalExperiments argues natural-experiment IV still needs an economic model to interpret the estimated parameter — a structuralist pushback on atheoretical “as-if random” identification.
• Weak and invalid instruments. Beyond exclusion, a weak first stage biases 2SLS toward OLS and breaks inference (Weak-Instruments, Murray2006-WeakAndInvalidInstruments); relevance is testable, validity is not.
• Shares vs. shocks in shift-share IV. The same Bartik instrument admits two incompatible identification stories: GoldsmithPinkhamEtAl2020-BartikInstruments reads it as a share-exogeneity design (identification from the exposure shares, shocks as weights), while BorusyakEtAl2022-QuasiExperimentalShiftShare reads it as a shock-exogeneity design (quasi-random shifters, endogenous shares fine given many uncorrelated shocks). Which story you adopt changes what must be defended and which diagnostics apply.
• Do shift-share standard errors lie? AdaoEtAl2019-ShiftShareDesigns shows that regions with similar sector shares have correlated residuals, so conventional/clustered standard errors over-reject dramatically (up to 55% in placebo) — a warning that identification is not the only fragile part of a Shift-Share-Instrument design.

Next

IV shares its potential-outcomes foundation with Foundations and its fuzzy-design and changes-in-changes machinery with DiD. Fuzzy RD is an IV problem at the cutoff — see RDD and CattaneoEtAl2020-RDDHandbook.