Staggered-adoption DiD and event study

Description

When units adopt a change at different times, estimate effects with event-study leads and lags and cohort-specific estimators that never use already-treated units as controls.

Products roll out to markets, accounts or teams at different dates, inviting difference-in-differences with two-way fixed effects. With staggered timing and effects that change over time, that pooled estimate is a weighted average of all 2x2 comparisons, including later-treated versus already-treated units, which can bias it or even flip its sign. The Goodman-Bacon decomposition shows each 2x2 estimate and its weight. The event study regresses outcomes on lead and lag indicators relative to each unit's adoption date, plotting pre-trends and dynamic effects; flat leads are reassuring but neither necessary nor sufficient for parallel trends. Heterogeneity-robust estimators (Callaway and Sant'Anna group-time ATTs, Sun and Abraham's interaction-weighted event study, stacked DiD) compare each cohort only with never- or not-yet-treated units and then aggregate with explicit weights.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Weeks

Steps

  1. Identify adoption cohorts by first treatment date and a clean control set of never-treated or not-yet-treated units.
  2. Estimate the static two-way fixed-effects DiD as a baseline and run the Goodman-Bacon decomposition to see each 2x2 comparison and its weight; check the weight on late-versus-early comparisons.
  3. Define event time relative to each unit's adoption and estimate leads and lags, omitting one pre-period as reference, with standard errors clustered by unit.
  4. Plot coefficients with 95% intervals; inspect leads for pre-trends and lags for dynamic effects, and repeat for first-stage outcomes such as exposure or take-up.
  5. Estimate cohort-by-period effects against clean controls (Callaway-Sant'Anna, Sun-Abraham or stacked DiD), optionally reweighting on covariates.
  6. Aggregate into overall or event-time summaries with explicit weights and compare with the naive estimate.

Sources

  • Causal Inference: The Mixtape (Cunningham, 2021)

Also known as

  • Event study with leads and lags
  • Dynamic DiD
  • Event-study DiD
  • Leads and lags plot
  • Goodman-Bacon decomposition
  • Bacon decomposition
  • Heterogeneity-robust DiD
  • Callaway-Sant'Anna estimator
  • Sun-Abraham interaction-weighted estimator
  • Stacked DiD
  • Group-time ATT

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference