Paper → Playable
Put a published methods result through its paces.
When the control is already treated.
Three equally sized groups, three periods. Early adopters start in period 1; late adopters in period 2; the third group is never treated. All untreated outcomes follow 10 + 2t. Early adopters’ initial effect is 2.
Read the method: Andrew Goodman-Bacon · Difference-in-Differences with Variation in Treatment Timing (2021; working paper 2018) ↗| Group | Period 0 | Period 1 | Period 2 |
|---|---|---|---|
| Early | 10 | 14 * | 20 * |
| Late | 10 | 12 | 16 * |
| Never | 10 | 12 | 14 |
What the paper establishes.
With staggered treatment timing, the conventional two-way fixed-effects estimator combines two-group, two-period comparisons. Comparisons using already-treated units can distort the target when treatment effects evolve.
Here the fixed-effects estimate is calculated by residualizing the treatment indicator by group and period, then applying OLS. The target shown is the equally weighted average effect across the three treated group-period cells. These are teaching data, not a replication of the paper’s empirical application. No standard errors are implied.
The never-treated group permits clean comparisons in this example. Real applications need a clearly defined target, defensible parallel trends and an estimator suited to treatment timing.
Read the method: Andrew Goodman-Bacon · Difference-in-Differences with Variation in Treatment Timing (2021; working paper 2018) ↗