VanillaSC – lost-autonomy triggers and secessionism

Estimator:

Vanilla Synthetic Control (VanillaSC) – mlsynth.VanillaSC

Source:

Schulte, F., Scantamburlo, M. and Ackren, M. (2026), “Lost autonomy triggers and the rise of secessionism,” European Political Science Review 18:175-193, DOI 10.1017/S175577392510026X.

Replication type:

Path A – reproduce the paper’s empirical finding on the authors’ data, and track the authors’ published synthetic-control series.

Status:

verified – mlsynth recovers the paper’s finding (a large post-trigger secessionist surge in both cases) and tracks the authors’ synthetic closely; the agreement with their specific package is close, not value-for-value.

What the paper does

Secessionist movements often surge suddenly. Schulte, Scantamburlo and Ackren (2026) argue that a “lost-autonomy trigger” – a dramatic event symbolising a loss of self-rule – can set off such a surge, and test this with synthetic control. They study two cases: the 2010 Spanish Constitutional Court decision that pared back Catalonia’s autonomy statute, and the 1994 economic shock (a major bank collapse and Denmark-imposed constraints) in the Faroe Islands. For each treated region they build a synthetic control from the other European autonomous regions and read the post-trigger gap in secessionist sentiment (av_sec1_all), with an in-space placebo (RMSPE-ratio) test for inference.

Their replication uses the SyntheticControlMethods Python package (Synth, pen="auto", n_optim=100) with covariate predictors – a penalized (Abadie-L’Hour) synthetic control with a V-optimized predictor match.

Method mapping

mlsynth’s canonical counterpart is VanillaSC with backend="outcome-only" – the well-posed convex problem, donor weights on the simplex minimising the pre-treatment outcome fit – and inference="placebo", the same in-space RMSPE-ratio test. The panel is the authors’ basedata/secession_autonomy.csv (13 autonomous regions, 1975-2021); each treated region is compared with the other twelve as donors, and treatment turns on at the trigger year (Catalonia 2010, Faroe Islands 1994).

The finding reproduces

Both methods deliver the paper’s headline – a large post-trigger surge in secessionist sentiment in both regions:

Quantity	VanillaSC (outcome-only)	Authors (SyntheticControlMethods)	correlation
Catalonia – post-2010 gap (ATT)	\(+28.0\)	\(+23.9\)	0.92
Faroe Islands – post-1994 gap (ATT)	\(+23.9\)	\(+26.8\)	0.75
Catalonia – pre-2010 fit RMSE	2.66	2.85
Faroe Islands – pre-1994 fit RMSE	4.92	5.05

Under the in-space placebo, Catalonia is the extreme case (RMSPE-ratio \(p \approx 0.077\), the smallest attainable with thirteen regions), while the Faroe result is weaker – consistent with the paper treating Catalonia as the cleaner test. The Catalonia synthetic leans on South Tyrol.

A note on tightness

This is a close reproduction, not a value-for-value one. The reference is a different implementation – SyntheticControlMethods optimises predictor (V) weights, auto-selects an Abadie-L’Hour penalty, and uses one hundred random restarts – so mlsynth’s deterministic outcome-only fit tracks the authors’ synthetic (correlation 0.75-0.92) and lands within about fifteen percent on the ATT, rather than matching digit for digit. mlsynth’s pre-treatment fit is in fact a touch tighter in both cases, which is expected: outcome-only minimises exactly the pre-period outcome error, whereas the authors’ penalized, covariate-matched fit trades a little of that for predictor balance and regularisation. When mlsynth is put on the same regularised footing (backend="penalized" or "mscmt" with the paper’s covariates) it lands right around the authors’ fit.

Because the cross-package agreement is approximate, this case appears on the Validation dashboard dashboard as a documented row – an honest cross-package comparison (same finding, different SC engine, about fifteen percent apart on the ATT) rather than a value-for-value exact or tight match.

Verification

The check is committed as the secession_scm benchmark case (benchmarks/cases/secession_scm.py), which runs VanillaSC on the shipped panel and asserts the deterministic ATT, pre-treatment RMSE, in-space placebo p-value, and the correlation with the authors’ published synthetic (captured under benchmarks/reference/secession_scm/). The fast offline unit test is mlsynth/tests/test_secession_replication.py.