Introduction

The methodology of causal inference has advanced rapidly since 2018. A cluster of papers — many published between 2018 and 2023 in the Journal of Econometrics, the American Economic Review, and the Review of Economic Studies — has fundamentally changed how applied researchers implement and interpret difference-in-differences designs. This issue's Results section takes stock of five of the most influential contributions.

Paper 1: Callaway and Sant'Anna (2021) — Staggered DiD

The Paper

Callaway and Sant'Anna(2021) — "Difference-in-Differences with Multiple Time Periods," Journal of Econometrics 225(2):200–230.

Main Finding

In settings with staggered treatment adoption and heterogeneous treatment effects, the classical TWFE estimator is invalid. The paper proposes group-time average treatment effects \(ATT(g,t)\) as the fundamental estimand, and a doubly robust estimator for each. The group-time ATTs can be aggregated into event-study, calendar-time, or simple overall estimates.

Identification Strategy

Identification rests on: (i) irreversibility of treatment; (ii) conditional parallel trends relative to a never-treated (or not-yet-treated) comparison group; and (iii) no anticipation. The doubly robust estimator is consistent if either the propensity score model or the outcome regression model is correctly specified.

Contribution

The paper resolved a longstanding confusion about what TWFE estimates in staggered settings, provided a rigorous and implementable alternative, and catalysed a wave of methodological and applied work. The companion R package did has been used in hundreds of applications across economics, public health, and political science, and the paper has become a standard reference for practitioners working with staggered treatment designs.

Paper 2: Goodman-Bacon (2021) — TWFE Decomposition

The Paper

Goodman-Bacon(2021) — "Difference-in-Differences with Variation in Treatment Timing," Journal of Econometrics 225(2):254–277.

Main Finding

The TWFE estimator in a staggered panel is a weighted average of all possible 2x2 DiD estimates: comparisons between each eventually-treated cohort and a "clean control" group across each pair of pre- and post-treatment periods. Some weights are negative when early-treated units serve as controls for later-treated units in periods when they are already under treatment. This negative weighting can produce sign reversals under sufficient treatment effect heterogeneity.

Identification Strategy

The decomposition is algebraic: no causal assumptions are needed. Goodman-Bacon shows that the TWFE coefficient equals: \[ \hat\delta^{TWFE} = \sum_{k \neq U} \sum_{l \neq k} \hat\mu_{kl} \hat\Delta^{DiD}_{kl} \] where the sum is over all pairs of cohorts \((k,l)\) and \(\hat\mu_{kl}\) are the regression weights, some of which may be negative.

Contribution

The decomposition theorem is both a diagnostic and a warning. Applied researchers can use the Goodman-Bacon decomposition (implemented in the bacondecomp package in R) to assess how much of their estimated effect comes from "clean" 2x2 comparisons and how much from "dirty" comparisons that use already-treated units as controls. The theorem provided the theoretical foundation for the rapid adoption of Callaway–Sant'Anna and Sun–Abraham estimators.

Paper 3: Rambachan and Roth (2023) — Honest Sensitivity Analysis

The Paper

Rambachan and Roth(2023) — "A More Credible Approach to Parallel Trends," Review of Economic Studies 90(5):2555–2591.

Main Finding

Pre-trend tests do not validate the parallel trends assumption for post-treatment periods. The paper proposes a framework for "honest" sensitivity analysis: the researcher specifies a set of plausible violations of parallel trends, and the paper derives confidence sets for the ATT that are uniformly valid over that set. The framework is implemented for two classes of restrictions: (i) the "smoothness" restriction, where post-trend violations cannot exceed a multiple \(\bar{M}\) of the maximum pre-trend violation; and (ii) the "sign" restriction, where violations are bounded and sign-constrained.

Identification Strategy

The key insight is to work with the sharp identified set for the ATT given a restriction on the trend violation. For a restriction that bounds the post-period counterfactual trend deviation by \(\delta^{\max}\), the identified set is a closed interval, and the confidence set is a union of intervals over the parameter set.

Contribution

The paper provides the first systematic framework for honest uncertainty quantification in DiD designs that accounts for potential parallel trends violations. It replaces the binary "pre-trend test passed / failed" with a continuous sensitivity parameter, enabling researchers to communicate the fragility of their conclusions in a principled way. The companion R package HonestDiD implements the framework and is now routinely required by some journals.

Paper 4: Ben-Michael, Feller, and Rothstein (2021) — Augmented Synthetic Control

The Paper

Ben-Michael et al.(2021) — "The Augmented Synthetic Control Method," Journal of the American Statistical Association 116(536):1789–1803.

Main Finding

The classical synthetic control estimator (Abadie et al.(2010)) can be biased when the pre-treatment fit is imperfect — when the treated unit lies outside the convex hull of the donor pool. The augmented synthetic control (ASCM) adds a ridge regression bias correction that removes the bias from imperfect fit under a linear factor model. The ASCM estimator achieves lower mean squared error than either SCM or simple regression adjustment, and it inherits the interpretability of the SCM weights.

Identification Strategy

The paper considers a linear factor model: \(Y_{it}(0) = \lambda_i' F_t + \varepsilon_{it}\), where \(\lambda_i\) are unit-specific factor loadings and \(F_t\) are common factors. Under this model, a convex combination of donor units can eliminate the factor-driven component of the outcome, and the ridge correction removes the remaining regression bias. The key tuning parameter (ridge penalty) is chosen by cross-validation on the pre-treatment fit.

Contribution

ASCM fills a gap between SCM (interpretable but potentially biased with imperfect fit) and OLS regression (low bias but potentially high variance). It performs especially well when the donor pool is large and the pre-treatment period is short — exactly the settings where classical SCM struggles. The companion R package augsynth extends the method to staggered adoption via multisynth.

Paper 5: Cengiz et al.\ (2019) — Bunching Estimator for Minimum Wages

The Paper

Cengiz et al.(2019) — "The Effect of Minimum Wages on Low-Wage Jobs," Quarterly Journal of Economics 134(3):1405–1454.

Main Finding

Using a bunching estimator applied to a near-universe of state-level minimum wage changes in the United States from 1979 to 2016, Cengiz et al.\ find that minimum wage increases raise wages at the bottom of the distribution without detectable disemployment. The number of jobs paying below the new minimum decreases, but this is fully offset by an increase in jobs paying at or just above the new minimum: there is no "missing" employment.

Identification Strategy

The paper constructs a "missing job" measure: for each minimum wage change, it counts the number of jobs in the wage range that is directly affected (between the old and new minimum) and compares this to the counterfactual number of jobs in that range, estimated from the distribution of wages in nearby (unaffected) wage bins. The identifying assumption is that the counterfactual wage distribution in the affected bin can be estimated from surrounding bins, using a difference-in-differences framework across wage bins and minimum wage changes.

Contribution

The bunching approach provides a novel way to aggregate evidence across many minimum wage changes, avoiding the comparison group problems that plague state-level DiD studies. By focusing on the wage distribution rather than total employment, it can detect effects that aggregate employment counts would miss, and it avoids the attenuation bias that affects employment-count measures when workers move up the wage distribution rather than out of employment. The paper has become a key reference in the minimum wage debate.

Conclusion

The five papers reviewed here collectively represent the state of the art in causal inference methodology as applied to policy questions. They share a commitment to rigorous identification, honest communication of assumptions, and the development of tools that allow researchers to learn from policy variation in administrative data. Each has spawned a software implementation that is now in routine use. The themes that emerge — heterogeneity, sensitivity to assumptions, and the value of aggregating across many quasi-experiments — define the frontier of the discipline.