Dynamic Difference-in-Differences
Dynamic Difference-in-Differences Estimator · Also known as: Dynamic DiD, Staggered DiD, Event-time DiD, Heterogeneous-timing DiD
Dynamic Difference-in-Differences extends the classic DiD framework to settings where units adopt treatment at different times. Rather than collapsing all variation into a single 2x2 comparison, it estimates group-time average treatment effects for each adoption cohort at each calendar period, then aggregates them into interpretable summaries of the causal effect over event time.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+18 more
When to use it
Use Dynamic DiD when treatment is adopted at different times across units (staggered rollout) and you suspect that treatment effects are heterogeneous across cohorts or change over event time. It is also appropriate when you want to formally test and visualise pre-trends, or when recovering event-time profiles matters more than a single average effect. It requires panel data with multiple pre- and post-treatment periods for at least some cohorts. Do not use it when all units are treated simultaneously (standard DiD suffices), when only one pre-treatment period exists per cohort (parallel trends are untestable), or when the panel is too short to trace dynamic paths meaningfully.
Strengths & limitations
- Avoids the contaminated weighted average produced by TWFE regression when treatment effects are heterogeneous across cohorts or event time.
- Recovers event-time profiles showing whether effects build up, stabilise, or fade — information standard DiD cannot provide.
- Provides a built-in formal pre-trend test rather than treating parallel-trends verification as an afterthought.
- Allows flexible aggregation into cohort-specific, calendar-time, or overall summaries matched to the research question.
- Implemented in widely-used, well-documented software packages such as R did and Stata csdid.
- Requires sufficient sample size within each adoption cohort; thin cohorts yield imprecise ATT(g,t) estimates that dominate aggregated results.
- Needs multiple pre-treatment periods per cohort to test parallel trends; a single pre-period leaves the key assumption untestable.
- Conceptually and computationally more demanding than TWFE, requiring deliberate choices about control group definition and aggregation weights.
- Identification still rests on conditional parallel trends; unobserved time-varying confounders that differ across cohorts invalidate the estimates.
- When no never-treated group exists and all units eventually adopt treatment, the not-yet-treated comparison group shrinks over time, increasing uncertainty in later event periods.
Frequently asked
How is Dynamic DiD different from a standard event-study regression?
A standard event-study adds leads and lags of a treatment dummy to TWFE regression but still uses already-treated units as controls for later-adopting cohorts, which can bias every coefficient when effects are heterogeneous. Dynamic DiD explicitly restricts each comparison to clean control groups — never-treated or not-yet-treated units — for each cohort-time cell, eliminating that contamination.
Do I need a never-treated group?
Not necessarily. Callaway and Sant'Anna allow the control group to be either never-treated units or not-yet-treated units. If every unit eventually adopts treatment, the not-yet-treated comparison is used, though this pool shrinks over time and can raise concerns if late adopters differ systematically from early ones.
How do I choose aggregation weights?
The choice should follow the research question. For an overall summary ATT, weight by cohort size and post-treatment exposure. For event-time dynamics, average ATT(g,t) over cohorts at each horizon. For group-specific effects, report cohort-level averages. The R did package and Stata csdid implement all three with a single argument.
What happens if treatment effects are homogeneous across cohorts?
When effects are truly constant across cohorts and event time, TWFE and Dynamic DiD produce the same estimate. The Dynamic DiD approach is not harmful in that case — it is simply more conservative — but the additional complexity is unnecessary. The TWFE critique only bites under heterogeneity.
Can Dynamic DiD be used with a binary outcome?
Yes. The ATT(g,t) framework is nonparametric and does not require a continuous outcome; it compares mean potential outcomes in differences. A linear probability interpretation applies, analogous to how standard DiD handles binary outcomes.
Sources
- Callaway, B., & Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200-230. DOI: 10.1016/j.jeconom.2020.12.001 ↗
- Sun, L., & Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics, 225(2), 175-199. DOI: 10.1016/j.jeconom.2020.09.006 ↗
How to cite this page
ScholarGate. (2026, June 3). Dynamic Difference-in-Differences Estimator. ScholarGate. https://scholargate.app/en/causal-inference/dynamic-difference-in-differences
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Difference-in-DifferencesEconometrics↔ compare
- Panel Data Difference-in-DifferencesCausal inference↔ compare
- Panel Fixed EffectsEconometrics↔ compare
- Synthetic Control MethodCausal inference↔ compare