⚡ TL;DR
The definitive 2026 synthesis of the post-2020 DiD revolution. Five of the field's principal methodologists pool the contributions of Goodman-Bacon (2021), Callaway-Sant'Anna (2021), Sun-Abraham (2021), Borusyak-Jaravel-Spiess (2024), de Chaisemartin-D'Haultfœuille (2020, 2023, 2024), Wooldridge (2023), Roth (2022), Rambachan-Roth (2023), Roth-Sant'Anna (2023), and others into one organizing framework. Likely to supersede Roth-Sant'Anna-Bilinski-Poe (2023) as the canonical reference for graduate teaching and applied work.
🧩 Setup & motivation
The paper organizes the modern DiD literature around three central questions: (1) What is the target parameter? — ATT, group-time ATT, dynamic event-time ATT, continuous-treatment dose-response, or aggregations thereof; (2) What are the identifying assumptions? — parallel trends, no anticipation, SUTVA, possibly conditional on covariates; (3) What estimator achieves the target under the assumptions? — and what diagnostics defend it.
The framework starts from the canonical 2×2 DiD and extends to: covariate adjustment (outcome regression, IPW, doubly robust); weights and aggregation; multiple periods; staggered treatments; continuous treatments; multi-shock designs; sensitivity analysis. Throughout, the authors emphasize matching the estimator to the target parameter — not picking an estimator based on what's popular.
📐 Main results
The unified framework
Two-by-two DiD identifies the ATT under parallel trends; staggered DiD requires either heterogeneity-robust estimators (CS, SA, BJS, dC-dH, Wooldridge ETWFE) OR strong homogeneity assumptions on TWFE — there is no free lunch. The paper systematizes when each estimator is appropriate, what target parameter it identifies, and what assumptions it requires.
Sensitivity analysis as default
Rambachan-Roth (2023) honest sensitivity bounds, Roth-Sant'Anna (2023) functional-form robustness, and Ghanem-Sant'Anna-Wüthrich (2025) selection-mechanism diagnostics are treated as part of standard practice, not optional robustness exercises. The framework integrates sensitivity bounds directly into the identification stage.
Software ecosystem
The paper provides a comprehensive table mapping each estimator to its R / Stata implementation: fixest::sunab, did::att_gt, didimputation::did_imputation, DIDmultiplegtDYN, etwfe::etwfe, HonestDiD, fwildclusterboot. Replication code for every example is on the GitHub repository.
Recommended workflow (paraphrased)
- Articulate target parameter (ATT? Group-time ATT? Dynamic event-time ATT?)
- State identifying assumptions explicitly
- Choose estimator that targets that parameter under those assumptions
- Run baseline + 3–4 robust estimators; report all
- Report Rambachan-Roth sensitivity bounds for headline coefficient
- Report Roth-Sant'Anna functional-form alternatives (levels vs logs)
- Honest discussion of remaining limitations
🛠️ Implications for practice
- This paper will likely become the canonical reference cited in DiD papers' methodology sections starting in 2026–2027.
- Graduate students should read this before applying DiD; faculty should adapt the framework's recommended workflow as the new default.
- If your paper does not pass the workflow checks here, expect reviewers to ask why.
- Pairs naturally with the Implementation Toolkit on this site (R code per step, package list, FAQ).
🧭 Where this sits in the broader DiD literature
Successor synthesis to Roth-Sant'Anna-Bilinski-Poe (2023) J Econometrics "What's Trending in DiD?", which itself synthesized the post-2020 methodology revolution. BCCGS 2026 will likely become the senior reference. Builds on every major paper from 2020 onward; integrates Ghanem-Sant'Anna-Wüthrich (2025) selection-mechanism perspective.
📥 Read the paper
- Local PDF (1.3 MB) — instant, no external request
- arXiv 2503.13323
- AEA Site
- GitHub replication