Using Synthetic Controls
Causal Question / Estimand
A survey/guide: when and how the synthetic control method credibly identifies the effect of an aggregate intervention on a single (or few) treated unit(s), and what data conditions that credibility.
Identification Strategy
Consolidates the synthetic-control framework: the counterfactual is a sparse, convex, non-extrapolating weighted average of donor units matched on a long pre-treatment outcome path. Grounds identification in a linear factor (interactive fixed effects) model — fitting many pre-periods aligns unobserved factor loadings, which is why a long, close pre-treatment fit (not merely matching covariate means) is what licenses the causal reading. Lays out feasibility conditions and inferential procedures (placebo/permutation tests, conformal inference).
Key Assumptions
Convex-Hull-Restriction, close fit over a long pre-treatment window, SUTVA (no interference, no anticipation embedded), No-Anticipation, donor pool of comparable, unaffected units, and sizeable signal-to-noise (effects large relative to volatility).
Threats to Validity
Poor pre-treatment fit (do not use SCM then); short pre-period; over-fitting noise; interpolation bias when the treated unit is outside the convex hull; donor contamination; specification/“researcher degrees of freedom” in predictor choice.
Setting / Data
n/a — methodological survey; recurring illustrations include California tobacco and German reunification.
Key Claims
- SCM is appropriate only with a long, close pre-treatment fit and a credible, unaffected donor pool — a bad fit is a reason not to use it.
- Transparency (explicit, sparse weights) and protection against extrapolation are SCM’s core advantages over regression and DiD.
- Inference must be design-based (placebo/permutation, conformal) because standard large-sample theory does not apply with one treated unit.
Connections
- Synthesizes: AbadieEtAl2010-SyntheticControlMethods, AbadieEtAl2015-ComparativePoliticsSCM
- Relates to DiD via factor models (see AtheyImbens2022-DesignBasedDiD) and to the broader program-evaluation toolkit (ImbensWooldridge2009-ProgramEvaluation).
- See also SCM
Citation
Abadie, A. (2021). Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects. Journal of Economic Literature, 59(2), 391–425.