How fragile is the difference-in-differences boom?
Difference-in-differences is the most-mentioned method on the equation ranking, and treatment effect heterogeneity is one of the fastest-rising keywords in economics. A first check finds that about three in ten recent DiD papers with searchable full text name a heterogeneity-robust estimator. How many staggered-adoption findings would survive one?
What prompted this
On the equation ranking, difference-in-differences is the most-mentioned method: 148 papers in the 30 days to 2026-10-23 (72 in economics journals, 53 in finance, 14 in management, 6 in accounting). Instrumental variables (112 papers), two-way fixed effects (69) and propensity score matching (60) follow. In economics, "treatment effect heterogeneity" had 14 papers in July–September 2026 against about 4 expected from its earlier share.
Two-way fixed-effects DiD can be badly biased when a policy starts at different times for different units and its effect varies. Newer estimators fix this: Callaway and Sant'Anna, Sun and Abraham, de Chaisemartin and D'Haultfœuille, and Borusyak, Jaravel and Spiess, with the Goodman-Bacon decomposition as a diagnostic.
First check
OpenAlex has searchable full text for 55 of the 148 DiD papers. 17 of those 55 (31%) name at least one of these estimators or the decomposition: 11 of 20 in finance, 3 of 21 in economics, 2 of 9 in management and 1 of 4 in accounting. The samples are small and most papers have no searchable full text, so this is a first signal, not a finding.
Questions
- Re-estimated with robust estimators, how often do recent staggered-adoption results change sign, size or significance?
- Which designs are most at risk: many adoption cohorts, long panels, or effects that build up over time?
- Do studies of Chinese pilot policies, a common setting in these journals, behave differently?
Data that could help
Replication packages where journals require them; for city pilot studies, the public city panels most of them use, such as the China City Statistical Yearbook.
Next step
Pick 20 recent staggered DiD papers with replication data and re-estimate each one with two robust estimators.