1. The counterfactual
Pick an estimator. The dashed line is the synthetic UK it produces; the shaded band after the referendum is the estimated shortfall.
2. Who is in the blend?
The unit weights. Four of the five recipes are nearly the same blend; only the augmented synthetic control is allowed to go negative.
3. Which quarters count?
Synthetic difference-in-differences also weights the pre-treatment quarters. Here almost all of that weight lands on one quarter, because log GDP behaves like a random walk.
4. The ladder, side by side
Every estimator's answer, against the 2.4% figure previously published for this dataset. Hover a bar for its note.
5. The fire drill
Move the treatment date back to a quarter when nothing happened, and every estimate is pure error. Lower is better. Switch the grading horizon to see why the published ranking of the three SDID variants does not survive.
6. Where the optimiser stopped
With no ridge penalty the objective is nearly flat, so the package's Frank–Wolfe solver halts on its iteration cap rather than at the optimum. The published estimate is the value it happens to be passing through.