Did the 2012 Kansas tax cut shrink the economy?
In 2012 Kansas enacted one of the largest state income-tax cuts in recent U.S. history — Governor Brownback called it "a real-live experiment." Because there is only one Kansas, the synthetic control method builds a counterfactual "synthetic Kansas" from a weighted blend of other states. The augmented version adds an outcome model that removes the bias left behind when the pre-2012 fit is imperfect.
This lab reproduces the tutorial's findings interactively, entirely client-side from a
precomputed results.json. Build the synthetic, watch augmentation deepen the
estimate, and test whether the effect could just be noise.
Kansas vs its synthetic control
Synthetic Kansas (blue) is a weighted blend of donor states built to match the real Kansas (orange) before 2012. After the dashed line, the gap between them is the estimated effect of the tax cut.
Before 2012 the two lines nearly coincide — a believable counterfactual. After 2012 Kansas slips below.
Build Synthetic Kansas
See the 7-state recipe and the classic-SCM gap with its conformal confidence band.
The Augmentation
Watch ridge regression estimate and subtract the bias — deepening the effect from −2.9% to −3.9%.
Inference
Four ways to ask "could this be noise?" — and why they can disagree.
Glossary (open a card if a term is unfamiliar)
Synthetic control
Augmentation (bias correction)
Pre-fit imbalance (L2)
ATT
Building synthetic Kansas with classic SCM
Classic SCM chooses non-negative weights that sum to one so the weighted donor history matches Kansas before 2012. The constraint forces most weights to exactly zero — the recipe is sparse and you can name it.
Who is in synthetic Kansas?
The donor weights. Hover a bar for the exact value. The other 42 states get exactly zero weight.
The classic-SCM gap, with conformal confidence band
Actual minus synthetic. Near zero before 2012 (good fit), persistently negative afterward. The shaded band is the pointwise conformal 95% interval.
Notice the dip near 2005–2006 before treatment — pure pre-fit imbalance that a convex blend cannot remove. That is the seed of bias the augmentation will fix.
Ridge augmentation: estimate the bias, then subtract it
When the pre-2012 fit is imperfect, the post-treatment gap mixes the real effect with bias. Ridge ASCM fits an outcome model, predicts the leftover imbalance, and subtracts it. A penalty λ — chosen by cross-validation — controls how far the weights may extrapolate.
Augmentation deepens the estimated effect
The classic-SCM gap (blue) and the Ridge-ASCM gap (teal). Both near zero before 2012; the ridge gap dips deeper afterward, because it has removed the bias from imperfect matching.
Choosing λ by cross-validation
Leave-one-pre-period-out CV error vs the penalty λ (log scale). The 1-SE rule picks the largest λ within one standard error of the minimum — the conservative choice.
The estimate grows as we de-bias
Average ATT across five specifications. More balancing → a more negative estimate and a better pre-fit. The un-augmented SCM number is the conservative one.
Inference: could the effect just be noise?
A synthetic-control estimate is a difference between two estimated curves built from one
treated unit, so significance is subtle. augsynth ships four tools; they share
the same −0.040 point estimate but can disagree on whether it is distinguishable from zero.
The same effect, four verdicts
Average ATT with each method's 95% interval (where it has one) or p-value. The dashed line is zero — an interval that crosses it is not significant at 5%.
Placebo test: is Kansas unusual?
Re-estimate the effect pretending each donor was treated, and compute its post/pre RMSPE ratio (how much it diverged after 2012, relative to its pre-2012 fit). Kansas (orange) should stand out if the effect is real. Each dot is one state.
Simulator: what makes an effect significant?
An illustrative treated-minus-synthetic gap: flat before treatment, shifted by the true
effect after. Move the sliders and watch the confidence interval widen or narrow and the
verdict flip at the 5% line.
(A teaching model — real augsynth uses conformal / jackknife, not this normal approximation.)
What this tab teaches
- Significance is not the point estimate. A clear −0.040 effect can still be borderline if the interval is wide.
- The methods probe different variation: conformal and jackknife+ over time; permutation and the leave-one-donor jackknife over units.
- They can disagree: jackknife+ excludes zero, while conformal (p = 0.066), permutation (p = 0.10) and the leave-one-donor jackknife are borderline.
- Be honest: report several methods and call the Kansas effect real but modest, not a knock-down result.