Basic Synthetic Control — Interactive Lab

A pedagogical companion to Basic Synthetic Control with R: The Basque Country Case Study ↗ Back to the post

What was the economic cost of conflict in the Basque Country?

The Basque Country experienced sustained terrorist activity starting in 1970. The honest counterfactual question — what would Basque GDP have been without the conflict? — cannot be answered by re-running history. The synthetic control method sidesteps this by building a weighted recipe of other Spanish regions whose pre-1970 economy looked indistinguishable from Basque, and using that synthetic stand-in as the missing counterfactual.

This app reproduces the post's three headline findings interactively. In four tabs you will see how a sparse two-donor recipe (85% Catalonia + 15% Madrid) is built, watch the actual–synthetic GDP gap open after 1970 and peak at −1.04 thousand USD per capita in 1989, and rank the Basque trace against placebo runs on the other 16 regions.

The headline finding in one picture

The synthetic Basque is built almost entirely from two of sixteen donor regions — Catalonia (85.1%) and Madrid (14.9%). This is what "sparse, convex combination" looks like when you plot the weights.

Catalonia and Madrid are the only two Spanish regions whose pre-1970 economies were comparably industrial, urban, and wealthy. A reader new to Spanish regional economics gets a one-line summary: Basque ≈ 85% Catalonia + 15% Madrid.

Tab 2

Donor Recipe

The full 16-region donor pool. See which regions were considered, and which got non-zero weight. Compare predictor balance side-by-side.

Tab 3

The Gap

Actual vs synthetic Basque, 1955–1997. Toggle the gap-only view to see the post-1970 divergence isolated from the levels.

Tab 4

Placebo Ranking

The full in-space placebo distribution. Trim by pre-treatment fit and see Basque rank 2 of 8 (pseudo p ≈ 0.25).

Glossary (open a card if a term is unfamiliar)

Synthetic control
A weighted average of donor (untreated) units designed to reproduce the treated unit's pre-treatment characteristics. The post-treatment path of the synthetic is the missing potential outcome.
Donor pool
The set of untreated units from which the synthetic is built. Here: 16 Spanish autonomous communities, excluding the Basque Country and the national aggregate.
Donor weights W
Non-negative weights summing to 1. Picked by the inner optimization to minimise the distance between treated and synthetic predictor vectors. The Basque solution gives 85.1% to Catalonia, 14.9% to Madrid, 0 to all others.
Predictor weights V
A diagonal matrix that tells the inner problem how much each predictor matters. Found by the outer loop, which minimises pre-treatment outcome MSPE.
Pre-treatment fit (MSPE)
Mean Squared Prediction Error of the synthetic on the pre-treatment outcome (gdpcap). Low MSPE pre-treatment is what justifies interpreting the post-treatment gap as causal.
ATT (gap)
The post-treatment difference between actual and synthetic outcomes. Headline estimate here is −0.58 thousand 1986 USD per capita averaged over 1970–1997.
In-space placebo
Run the algorithm pretending each donor region was the treated unit. Compare its post/pre MSPE ratio to Basque's. The Basque ratio of 60.1 ranks 7 of 17 (full) or 2 of 8 (trimmed by pre-fit comparability).
MSPE ratio (post / pre)
A normalised statistic that lets placebo regions with different fit qualities be compared. Larger ratio = larger relative gap. Basque: 60.1. Catalonia: 64.7.

The donor recipe — who builds the synthetic Basque?

The optimization searches the 16-region donor pool for the convex combination whose pre-treatment predictors look most like Basque. The result is a sparse two-donor solution. Below: the optimised weights for all 16 donor regions (sorted), plus the predictor balance that explains why these two were picked.

Optimal donor weights

Steel blue = dominant donor (Catalonia). Orange = secondary donor (Madrid). All other regions are dimmed because their optimised weights are exactly zero.

Predictor balance — why Catalonia + Madrid wins

For each of the 13 matched predictors, this table compares the treated Basque value, the synthetic Basque value (from the optimised W), and the simple average across the full donor pool. Click a row to highlight.

Look at special.gdpcap.1960.1969: treated 5.285 vs synthetic 5.271 (sample mean: 3.581). The match on pre-1970 GDP is essentially perfect — that is why the post-1970 divergence is informative. Look at special.sec.industry.1961.1969: treated 45.1% industry share vs synthetic 37.6% vs sample mean only 22.4%. The optimiser correctly identified Basque as highly industrial and shifted weight onto donors that share that profile.

What to look for

  • The two-donor sparsity is a feature, not a bug. When only a few donors closely resemble the treated unit, the optimiser correctly concentrates weight on them. The other 14 get exactly zero — no contribution.
  • The match is excellent on outcome-relevant predictors (pre-1970 gdpcap, education shares within 1–4 points of each other) and acceptable on industrial structure. The residual gap on population density (treated 247 vs synthetic 196) is the largest — but density is a static control rather than a year-to-year outcome predictor.
  • Reproducible recipe. The same weights drop out of the BFGS optimiser every time you run the post's analysis.R — there is no stochastic step here.

Actual Basque vs synthetic Basque, 1955–1997

These are the real numbers from the post's gap_series.csv. Pre-1970 (grey shaded), the two paths should be indistinguishable — that is the point of the construction. Post-1970, any divergence is the estimated economic cost of conflict, expressed in thousand 1986 USD per capita.

View

Average ATT (1970–1997)
−0.580
thousand 1986 USD per capita
Peak gap (1989)
−1.036
single-year maximum deficit
End-of-window gap (1997)
−0.828
deficit at last observation
Pre-1970 fit MSPE
0.0082
near-perfect match before treatment

What to look for

  • Pre-1970 the two lines are nearly indistinguishable. Both rise from about 3.8 to 6.2 thousand 1986 USD across the 15 pre-treatment years. The MSPE is 0.0082 — the synthetic is doing what it was constructed to do.
  • Starting around 1972 the lines diverge. By 1989 the gap reaches its largest single-year deficit of −1.04 thousand USD per capita. This corresponds to roughly an 8% income shortfall against the no-conflict counterfactual.
  • The recovery is partial. After 1989 the gap narrows somewhat (the synthetic also moderates) but stays around −0.8 thousand USD all the way to 1997. There is no return to the pre-treatment pattern of overlapping paths.

In-space placebo distribution — is the Basque gap unusual?

The same algorithm is re-run with each of the other 16 regions as the "treated" unit. If terrorism had no causal effect on Basque GDP, then the Basque post/pre MSPE ratio should be unremarkable in this distribution. The chart below shows where it actually sits.

Filter

Basque ratio (post / pre MSPE)
60.1
pre-MSPE 0.0082, post-MSPE 0.493
Rank (trimmed)
2 of 8
pseudo p ≈ 0.250
Rank (full, untrimmed)
7 of 17
pseudo p ≈ 0.412
Catalonia ratio (top placebo)
64.7
the donor that builds 85% of Basque

How to read this

  • Toggle to "Full" view to see why trimming matters. Andalucia, Asturias, and Navarra all have astronomical ratios (904, 549, 388) — not because they had real shocks, but because their pre-1970 fit was so good that even a moderate post-period gap divides into a huge ratio. They are not comparable to Basque.
  • Switch back to "Trimmed." Among the 8 regions whose pre-MSPE is within a factor of 5 of Basque's, Basque ranks 2 of 8 (pseudo p = 1/4 = 0.250). The first-place region is Catalonia, which contributes 85% of the synthetic Basque recipe — when a treated unit is built mostly from one donor, that donor is structurally a likely placebo because its own synthetic struggles to find a substitute.
  • The honest takeaway from this tab: the visual signal in Tab 3 is suggestive of a real economic cost, but the formal inference is hampered by the small donor pool (smallest pseudo p possible in the trimmed set = 1/8 ≈ 0.125, so even the maximum-rank placebo would not clear conventional significance thresholds).

Connecting back to the previous tabs

The placebo evidence and the donor-weight structure (Tab 2) should be read together, not separately. The same Catalonia that builds 85% of synthetic Basque also produces the largest single-region placebo gap. That is not a coincidence — it is a structural feature of the method when only a few donors closely resemble the treated unit. The Basque finding survives this scrutiny but is not bulletproof; the post's recommendation is to consult scpi (Cattaneo, Feng & Titiunik 2021) for prediction intervals when frequentist inference is essential.