Observational Causal Inference

The foundational paper introducing propensity scores for causal inference in observational studies.

Best practices for matching methods and the MatchIt software implementation.

The double ML framework using cross-fitting to obtain valid inference with ML first-stage estimation.

Introduced entropy balancing, which achieves exact covariate balance through maximum entropy reweighting—eliminating iterative propensity score model searching.

Accessible exposition of the augmented IPW estimator, consistent if either propensity score or outcome model is correct—the foundational 'doubly robust' property.

Foundational theoretical paper deriving semiparametric efficiency bounds and introducing the AIPW estimator class underlying all modern doubly robust methods.

Comprehensive review of modern DiD methods including staggered adoption and heterogeneous effects.

Decomposes two-way fixed effects estimators and reveals issues with staggered DiD designs.

Group-time average treatment effects and aggregation methods for staggered DiD.

Demonstrates TWFE regressions estimate weighted sums of ATEs with potentially negative weights—proposing the DIDM estimator as solution.

Shows TWFE event-study coefficients are contaminated by effects from other periods, proposing an interaction-weighted estimator.

Formal sensitivity analysis for parallel trends violations—implemented in the widely-used HonestDiD package.

The foundational synthetic control paper with the California tobacco application.

Practical guidance on when and how to apply synthetic control methods.

Combines synthetic control and DiD for improved inference in panel data settings.

Bridges matrix completion/ML with synthetic control using nuclear norm regularization—handles staggered adoption and outperforms traditional SC.

Extends synthetic control to settings where perfect pre-treatment fit is infeasible using ridge regression to de-bias estimates.

Essential theoretical foundation for statistical inference in SC applications, extending permutation tests and constructing proper confidence sets.

Non-parametric algorithm for detecting treatment effects over time using synthetic controls with ML methods (Random Forest, Lasso, etc.) without assuming correct model specification.

The LATE framework for interpreting IV estimates as effects on compliers.

Defines the assumptions needed for IV and connects to potential outcomes framework.

Foundational framework for detecting weak instruments—the origin of the 'first-stage F > 10' rule now required in all IV applications.

Modern econometric framework for Bartik instruments—identification follows from quasi-random shock assignment rather than exogenous shares.

Develops nonparametric tests for exclusion and monotonicity in examiner IV designs—essential for criminal justice, disability, and immigration research.

Comprehensive guide to RDD identification, estimation, and practical implementation.

Modern implementation guide with rdrobust software for sharp and fuzzy RDD.

Introduces the canonical density discontinuity test for detecting manipulation at the cutoff—now a required falsification check in all RDD work.

Derives the MSE-optimal bandwidth for local linear RD estimation—the first principled approach to bandwidth selection.

Shows MSE-optimal bandwidths yield invalid conventional CIs and develops bias-corrected robust inference—foundation for the rdrobust package.

Causal forests for estimating conditional average treatment effects with valid confidence intervals.

Framework for using any ML method to find heterogeneous effects with valid inference.

Taxonomy of metalearners (S-learner, T-learner, X-learner) for CATE estimation.

Introduces the R-learner framework achieving quasi-oracle efficiency—matching error bounds of an oracle knowing nuisance components.

Establishes model-free oracle inequalities for the DR-learner—doubly robust CATE estimation achieves faster convergence rates.

Theoretical foundations for learning optimal treatment policies from observational data using doubly robust scores.

The E-value for quantifying sensitivity to unmeasured confounding.

Modern sensitivity analysis framework with intuitive benchmarking against observed covariates.

Foundational paper establishing the partial identification paradigm—showing what can be learned under minimal assumptions when point identification fails.

Developed the 'Lee bounds' trimming procedure for handling attrition under monotonicity—now a standard robustness check for differential selection.

Shows how to jointly use coefficient movements and R-squared changes to bound omitted variable bias—the workhorse sensitivity analysis.

Pioneered the insight that selection on observables guides selection on unobservables—framework for assessing confounding needed to explain away effects.