Open-Source Econometrics · ERC Really Credible

Which Difference-in-Differences
Estimator Should I Use?

There is no universally best Difference-in-Differences estimator. The right estimator depends on the design of your study and the effect you want to estimate.

Is treatment binary or nonbinary? Absorbing, meaning that once treated one always remains treated, or reversible? Do treated groups adopt at different dates? Can treatment change more than once? Is it plausible that lagged treatments still affect the current outcome? Are there untreated or not-yet-treated groups, stayers, or quasi-untreated observations? This guide first helps you characterize your research design, and then explains which modern DiD estimators are compatible with it.

Stata users can also run whichDIDshouldIuse for an automatic characterization of their design — see below.

Further questions? Ask them to ChatDID.

Compare

Identify your design, then pick an estimator

The table below lists the designs reviewed in the textbook Causal Inference with Differences-in-Differences: Credible Answers to Hard Questions by de Chaisemartin and D'Haultfœuille. For each design, it lists possible estimators, restricting attention to heterogeneity-robust estimators that only rely on no-anticipation and parallel-trends assumptions. Most commands and packages listed below can be used to conduct pre-trend tests of those assumptions.

NameDefinition & examplesTreatment patternEstimatorsCommands & packagesFurther reading
Canonical DID design

Treatment is binary, absorbing, there are never-treated units, and there is no variation in treatment timing: all treated units start getting treated at the same date.

Examples

Cholera is water borne. Some London neighborhoods change their water supply to less contaminated (upstream) water, other neighborhoods do not; cholera outbreaks decrease more in the first group. Snow (1856)

 Without control variables
  • TWFE regression
  • Event-study TWFE regression
With control variables
  • Heterogeneity-robust estimators, see below
 Without controls
reg / xtreg / reghdfe
With controls

See below.

 Chapters 3 and 4 of the textbookBook & slides
Staggered adoption design

Treatment is binary, absorbing, and there is variation in treatment timing.

Examples

Free trade agreements. Country pairs sign FTAs at different dates and never repeal them; FTAs raise bilateral trade. Baier & Bergstrand (2007)

Unilateral divorce. Introduced at different dates across Western Europe, never repealed; it raises divorce rates. Kneip & Bauer (2009)
  • Heterogeneity-robust DID estimators using the never-treated, the not-yet-treated or the last-treated as the control group
  • Imputation estimators
  • Should I use a heterogeneity-robust DID estimator or an imputation estimator?
    • In Chapter 6 of the textbook, we conduct a meta-analysis of 12 political science articles using a staggered adoption design. Heterogeneity-robust DID estimators and imputation estimators are always close to each other, except in the two papers where pre-trend tests are rejected. We recommend using DID estimators using not-yet-treated as controls: they are more transparent than imputation estimators, and in our meta-analysis they are slightly less noisy than imputation estimators.
Heterogeneity-robust DID estimators
eventstudyinteract (never- or last-treated)
csdid (never- or not-yet-treated)
did_multiplegt_dyn (never- or not-yet-treated)
Imputation estimators
did_imputation
did2s
fect
jwdid (etwfe in R)
Ch. 6 of the textbookde Chaisemartin & D'Haultfœuille (2020)de Chaisemartin & D'Haultfœuille (2026)Callaway & Sant'Anna (2021)Sun & Abraham (2021)Gardner (2021)Liu, Wang & Xu (2024)Borusyak, Jaravel & Spiess (2024)Wooldridge (2025)
Heterogeneous adoption design

All units are untreated at the start of the panel, all treated units become treated at the same date, and treated units receive heterogeneous treatment doses.

Examples

Minimum wage. German districts more exposed to an increase of the national minimum wage saw larger wage gains and no employment losses. Dustmann et al. (2022)

Cancer treatment. US cancer programs adopted a new chemotherapy at different rates; short-term mortality decreased more among high adopters. Melamed et al. (2021)

 Are there never-treated units?
  • Yes: a DID estimator using the never-treated as controls, scaled by the average dose received by treated units.
  • No: are there quasi-untreated units?
    • Yes: a DID estimator using the quasi-untreated as controls, scaled by the average dose received by treated units.
    • No: a DID estimator using the least-treated as controls, scaled by the average dose received by treated units. NB: this estimator relies on a homogeneous effect assumption.
 Never-treated units: yes
ivregress 2sls Δy (d = 1(d>0))
Never-treated units: no (with or without quasi-untreated units)
did_had

did_had runs a test of the null hypothesis that your design has quasi-untreated units.

 Ch. 7 of the textbookde Chaisemartin et al. (2026)
Staggered first switch design

Designs where groups experience their first treatment change at different points in time. These designs nest staggered adoption designs. But staggered first switch designs also encompass more general designs, where the treatment may be nonbinary and/or non-absorbing.

Examples

Democratic governors. US states move in and out of Democratic governorship (binary, non-absorbing); effect on policy liberalism. Caughey, Xu & Warshaw (2017)

Business taxes. German municipalities start at different tax rates and change them repeatedly; effect on wages. Fuest, Peichl & Siegloch (2018)

 Estimators allowing for effects of lagged treatments
  • DID estimator comparing first-time switchers (units whose treatment changes for the first time) to not-yet switchers (units whose treatment has not changed yet), controlling for units' treatment at period one
If lagged effects can plausibly be ruled out
  • DID comparing the t−1 to t outcome evolutions of units whose treatment switches and does not switch between t−1 and t, controlling for units' treatment at t−1
  • Imputation estimators, if the treatment is binary
 Estimators allowing for effects of lagged treatments
did_multiplegt_dyn
If lagged effects can be ruled out, at least beyond a pre-specified number of lags
did_multiplegt_stat (DID estimator)
did_imputation
did2s
fect

did_imputation, did2s and fect are imputation estimators, valid if the treatment is binary.

 Ch. 8 of the textbookde Chaisemartin & D'Haultfœuille (2020)de Chaisemartin & D'Haultfœuille (2026)de Chaisemartin et al. (2022)Gardner (2021)Liu, Wang & Xu (2024)Borusyak, Jaravel & Spiess (2024)
Other designs

Designs that do not fall into any of the designs above. For instance, the treatment of every unit could change between every pair of time periods.

Example

Weather. Rainfall or temperature differs across locations and changes every year, e.g. to study agricultural yields. Deschênes & Greenstone (2007)

 No heterogeneity-robust estimator available for these designs. NA Section 8.5.2 of the textbookde Chaisemartin et al. (2024)

Note: designs with several treatments

The table above only considers designs with a single treatment. Sometimes researchers want to estimate the effects of several treatments.

  • Definition. Units may receive K ≥ 2 treatments, or a treatment with several categories. This framework also allows accounting for spillover effects, by treating exposure to other units' treatment as an additional treatment.
  • Examples.
    • Medical and recreational marijuana laws in the US, whose effects one may want to estimate separately. So far, states have always passed the former before the latter (see, e.g., Meinhofer, Witman, Hinde and Simon, 2021).
    • Voter ID laws in the US, where less strict laws are typically passed before stricter ones (see Cantoni and Pons, 2021).
    • Anti-deforestation policies in the Amazon rainforest, where plots of land are first put into a concession, and then some concessions get certified (see Rico-Straffon, Wang, Panlasigui, Loucks, Swenson and Pfaff, 2023).
  • Estimation.
    • TWFE regressions: the twowayfeweights Stata and R commands compute the weights attached to TWFE regressions with several treatments, using the other_treatments option.
    • If lagged effects can be ruled out: heterogeneity-robust estimators of the effect of each treatment can be computed with did_multiplegt_stat.
    • Dynamic effects with sequential treatments: the additional effect of the second treatment can be estimated with did_multiplegt_dyn, restricting the sample to the (g, t)s that have received the first treatment, and including the adoption date of the first treatment in the trends_nonparam option.

See Section 8.3.5.4 of the textbook for more details.

Identify your design automatically in Stata

Run whichDIDshouldIuse on your data for an automatic characterization of your design and a menu of estimators you can use given your design.

This diagnostic helps classify your design; it is not a substitute for assessing the identifying assumptions underlying each estimator.

Stata ssc install whichDIDshouldIuse
Stata ssc install distinct
Stata whichDIDshouldIuse Y G T D

Requires distinct (SSC). Not yet on SSC for your version? Check the GitHub repository.

References

  1. Baier, S. L., & Bergstrand, J. H. (2007). "Do Free Trade Agreements Actually Increase Members' International Trade?" Journal of International Economics, 71(1), 72–95.
  2. Borusyak, K., Jaravel, X., & Spiess, J. (2024). "Revisiting Event Study Designs: Robust and Efficient Estimation." Review of Economic Studies, 91(6), 3253–3285. DOI
  3. Callaway, B., & Sant'Anna, P. H. C. (2021). "Difference-in-Differences with Multiple Time Periods." Journal of Econometrics, 225(2), 200–230. DOI
  4. Cantoni, E., & Pons, V. (2021). "Strict ID Laws Don't Stop Voters: Evidence from a U.S. Nationwide Panel, 2008–2018." Quarterly Journal of Economics, 136(4), 2615–2660.
  5. Caughey, D., Xu, Y., & Warshaw, C. (2017). "Incremental Democracy: The Policy Effects of Partisan Control of State Government." Journal of Politics, 79(4), 1342–1358.
  6. de Chaisemartin, C., & D'Haultfœuille, X. (2020). "Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects." American Economic Review, 110(9), 2964–2996. DOI
  7. de Chaisemartin, C., D'Haultfœuille, X., Pasquier, F., & Vazquez-Bare, G. (2022). "Difference-in-Differences for Continuous Treatments and Instruments with Stayers." Working paper, SSRN 4011782. SSRN
  8. de Chaisemartin, C., & D'Haultfœuille, X. (2023). "Two-Way Fixed Effects and Differences-in-Differences Estimators with Several Treatments." Journal of Econometrics, 236(2), 105480.
  9. de Chaisemartin, C., D'Haultfœuille, X., & Vazquez-Bare, G. (2024). "Difference-in-Difference Estimators with Continuous Treatments and No Stayers." AEA Papers and Proceedings, 114, 610–613.
  10. de Chaisemartin, C., Ciccia, D., D'Haultfœuille, X., & Knau, F. (2026). "Difference-in-Differences Estimators When No Unit Remains Untreated." Working paper, arXiv:2405.04465. arXiv
  11. de Chaisemartin, C., & D'Haultfœuille, X. (2026). "Difference-in-Differences Estimators of Intertemporal Treatment Effects." Review of Economics and Statistics, 108(4), 863–880. Journal page
  12. Deschênes, O., & Greenstone, M. (2007). "The Economic Impacts of Climate Change: Evidence from Agricultural Output and Random Fluctuations in Weather." American Economic Review, 97(1), 354–385.
  13. Dustmann, C., Lindner, A., Schönberg, U., Umkehrer, M., & vom Berge, P. (2022). "Reallocation Effects of the Minimum Wage." Quarterly Journal of Economics, 137(1), 267–328.
  14. Fuest, C., Peichl, A., & Siegloch, S. (2018). "Do Higher Corporate Taxes Reduce Wages? Micro Evidence from Germany." American Economic Review, 108(2), 393–418.
  15. Gardner, J. (2021). "Two-Stage Differences in Differences." Working paper. arXiv
  16. Goodman-Bacon, A. (2021). "Difference-in-Differences with Variation in Treatment Timing." Journal of Econometrics, 225(2), 254–277. DOI
  17. Kneip, T., & Bauer, G. (2009). "Did Unilateral Divorce Laws Raise Divorce Rates in Western Europe?" Journal of Marriage and Family, 71(3), 592–607.
  18. Liu, L., Wang, Y., & Xu, Y. (2024). "A Practical Guide to Counterfactual Estimators for Causal Inference with Time-Series Cross-Sectional Data." American Journal of Political Science, 68(1), 160–176. DOI
  19. Melamed, A., et al. (2021). JAMA Oncology. DOI
  20. Meinhofer, A., Witman, A. E., Hinde, J. M., & Simon, K. (2021). "Marijuana Liberalization Policies and Perinatal Health." Journal of Health Economics, 80, 102537.
  21. Rico-Straffon, J., Wang, Z., Panlasigui, S., Loucks, C. J., Swenson, J., & Pfaff, A. (2023). "Forest Concessions and Eco-Certifications in the Peruvian Amazon: Deforestation Impacts of Logging Rights and Logging Restrictions." Journal of Environmental Economics and Management, 118, 102780.
  22. Snow, J. (1856). "On the Mode of Communication of Cholera." Edinburgh Medical Journal, 1(7), 668–670.
  23. Sun, L., & Abraham, S. (2021). "Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects." Journal of Econometrics, 225(2), 175–199. DOI
  24. Wooldridge, J. M. (2025). "Two-Way Fixed Effects, the Two-Way Mundlak Regression, and Difference-in-Differences Estimators." Empirical Economics, 69(5), 2545–2587. DOI