Abstract
Difference-in-differences (DiD) is the workhorse of empirical economics when a policy hits some places and not others. The classic two-way fixed-effects regression is still taught as if it always recovers a single average treatment effect. When units adopt treatment in different years, that regression is a weighted average of many 2×2 comparisons—including already-treated units used as controls for later adopters. Goodman-Bacon showed that some of those weights can be negative, so the coefficient can have the opposite sign of every unit-level effect. This paper explains the problem without matrix algebra, reviews Callaway–Sant’Anna and related estimators, and argues that student projects using state-by-year policies (minimum wages, school-mask rules, sports betting legalization) should not report a single TWFE coefficient as if it were a fact about the world.
The Comparison You Did Not Know You Were Making
In a simple DiD, untreated units trace the counterfactual path of treated units. Staggered adoption breaks the simplicity. A state treated in 2014 is a control for a state treated in 2018, even though it is no longer untreated. If treatment effects change over time, that comparison is contaminated (Goodman-Bacon). Negative weights arise when already-treated units with growing effects are used as baselines, pulling the average in the wrong direction.
This is not a rare pathology. Many education, labor, and health papers use staggered rollouts. Roth et al. survey how quickly the applied literature has had to rewrite itself. The niche for students is that “I controlled for state and year fixed effects” is no longer a complete methods sentence.
Estimators That Respect Cohorts
Callaway and Sant’Anna estimate group-time average treatment effects using only not-yet-treated or never-treated units as controls, then aggregate with explicit weights (Callaway and Sant’Anna). Sun and Abraham propose an interaction-weighted estimator that avoids forbidden comparisons. Imputation approaches (Borusyak et al.) fill in untreated potential outcomes rather than stacking contaminated 2×2s. None of these is magic: they still need parallel trends, and they still need enough never-treated or late-treated units.
Event-study plots remain essential. If pre-trends slope, the design is already in trouble, staggered or not. New estimators can be stacked on a bad design and still lie. The upgrade is honesty about which comparisons enter the average.
A Rule for Student Empirical Work
If your policy turns on in the same year everywhere, classic DiD may still be the right teaching example. If it dribbles out across years, report a cohort-aware estimator and the TWFE number side by side. If they disagree, believe the one that does not use treated units as controls—or admit that the design cannot answer the question.
Economics at the high-school and early-college level should include this caveat because it is now central, not obscure. Staggered DiD is a niche of econometric hygiene that changes how we read headlines about “the effect of the policy.”
Why Two-Way Fixed Effects Break
Classical difference-in-differences with two groups and two periods estimates a single treatment effect under parallel trends. Staggered adoption with two-way fixed effects (TWFE) instead averages many 2×2 comparisons, some of which use already-treated units as controls. When effects grow over time, those comparisons can be negatively weighted, and the TWFE coefficient can even have the wrong sign (Goodman-Bacon; Callaway and Sant’Anna).
That is not a niche complaint about econometric fashion. It means a student who regresses test scores on “years since a laptop program started,” with school and year dummies, may not be estimating the effect they think they are. The niche literacy is to ask which comparisons the software is actually averaging.
Estimators Students Will See in New Papers
Callaway and Sant’Anna group-time average treatment effects, Sun and Abraham cohort-event-study interactions, and imputation estimators that fit untreated potential outcomes first are now standard replies to TWFE (Callaway and Sant’Anna; Sun and Abraham; Borusyak et al.). They require stating an assumption about which not-yet-treated or never-treated units are valid counterfactuals, and they often need clustered uncertainty.
Parallel trends is still an assumption, not a theorem. Event-study plots of pre-trends help but can be underpowered. A mid-size paper should treat diagnostics as part of the result, not as an appendix dumped after a star-studded coefficient.
Conclusion
Staggered difference-in-differences is a niche methodological topic with outsized consequences for education, labor, and health policy papers. The complex idea is weighting: who is compared to whom, and when.
Student researchers should not be expected to invent a new estimator. They should be expected not to copy a 2010 two-way fixed-effects specification uncritically in 2026. That standard is how a high-school journal can still be technically serious.
Works Cited
Goodman-Bacon, Andrew. “Difference-in-Differences with Variation in Treatment Timing.” Journal of Econometrics, vol. 225, no. 2, 2021, pp. 254–77.
Callaway, Brantly, and Pedro H. C. Sant’Anna. “Difference-in-Differences with Multiple Time Periods.” Journal of Econometrics, vol. 225, no. 2, 2021, pp. 200–30.
Sun, Liyang, and Sarah Abraham. “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects.” Journal of Econometrics, vol. 225, no. 2, 2021, pp. 175–99.
Roth, Jonathan, et al. “What’s Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature.” Journal of Econometrics, vol. 235, no. 2, 2023, pp. 2218–44.
Borusyak, Kirill, Xavier Jaravel, and Jann Spiess. “Revisiting Event-Study Designs: Robust and Efficient Estimation.” Review of Economic Studies, vol. 91, no. 6, 2024, pp. 3253–85.
Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. “How Much Should We Trust Differences-in-Differences Estimates?” Quarterly Journal of Economics, vol. 119, no. 1, 2004, pp. 249–75.
de Chaisemartin, Clément, and Xavier D’Haultfœuille. “Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects.” American Economic Review, vol. 110, no. 9, 2020, pp. 2964–96.