Synthetic Control Method

Synthetic control method (abbreviated SCM) is a data-driven quasi-experimental causal inference method in epidemiology, biostatistics, and public health policy evaluation developed by Alberto Abadie and Javier Gardeazabal (2003) and formalized by Alberto Abadie, Alexis Diamond, and Jens Hainmueller (2010, 2015) to evaluate aggregate interventions enacted in a single treated unit—such as a state, nation, municipality, or hospital system—when no single untreated comparison jurisdiction exhibits a parallel pre-intervention trend. Rather than arbitrarily selecting a single neighboring state or taking a simple unweighted average of all untreated jurisdictions (as in traditional difference-in-differences), the synthetic control method constructs an artificial counterfactual unit—the synthetic control—as a convex weighted average of untreated units drawn from a candidate donor pool. The non-negative weights (w_j ≥ 0, ∑ w_j = 1) are optimized algorithmically so that the synthetic control closely matches the treated jurisdiction's pre-intervention outcome trajectory and baseline prognostic covariates over T₀ pre-policy periods. The causal impact of the intervention at each post-policy time point (t > T₀) is then estimated by the difference between the observed outcome of the treated unit and the concurrent outcome of its synthetic counterpart.

Why Single-Jurisdiction Comparisons Fail in Public Health

Public health interventions are frequently enacted by a single pioneering jurisdiction: California's 1988 Proposition 99 tobacco tax and clean-air initiative, Massachusetts's 2006 health insurance reform, Brazil's national introduction of conjugate pneumococcal vaccine, or England's teenage pregnancy strategy. When epidemiologists attempt to evaluate such policies using standard difference-in-differences or controlled interrupted time series, two structural difficulties arise:

1. Violation of Parallel Pre-Trends: A pioneering state or country often adopts a policy precisely because its baseline disease burden, demographic structure, or secular health trajectory differs from the national average, violating the parallel-trends assumption.
2. Subjective Control Selection and Extrapolation: Selecting a single "similar" neighboring state invites researcher degrees of freedom, whereas linear regression across all untreated states relies on unconstrained negative and > 1 regression weights that extrapolate outside the convex hull of the observed data.

As proved by Abadie, Diamond, and Hainmueller (2010) using a linear factor model with time-varying unobserved confounders (Y_jt(0) = δ_t + θ_t^TZ_j + λ_t^Tμ_j + ε_jt), if a weighted combination of donor units matches both the observed covariates Z₁ and a sufficiently long sequence of pre-intervention outcomes (Y_1,1, ..., Y_1,T0), the synthetic control implicitly matches the treated unit's unobserved time-varying factor loadings (μ₁ ≈ ∑ w_j μ_j), neutralizing unmeasured confounding that changes over time.

Donor Pool, Convex Weight Constraints, and Counterfactual Estimator

Suppose we observe J + 1 aggregate jurisdictions indexed by j = 1, 2, ..., J + 1 across T calendar periods (t = 1, ..., T₀, T₀ + 1, ..., T). Without loss of generality, let unit j = 1 be the single treated jurisdiction exposed to the public health intervention from time T₀ + 1 onward, while the remaining J jurisdictions (j = 2, ..., J + 1) remain untreated throughout the study window and constitute the donor pool.

A synthetic control is defined by a J × 1 weight vector W = (w₂, w₃, ..., w_J+1)^T satisfying two convexity constraints:

1. Non-Negativity: w_j ≥ 0 for all j ∈ {2, ..., J + 1}
2. Unit Sum (No Intercept / No Extrapolation): ∑_j=2^J+1 w_j = 1

Restricting weights to the unit simplex guarantees that the synthetic control is an interpolation inside the convex hull of real donor jurisdictions and naturally induces sparsity: most donor units receive an exact weight of w_j = 0, leaving a transparent combination of 3 to 6 identifiable donor regions whose percentage contributions sum to 100% (for example, "Synthetic California" in Abadie et al.'s Proposition 99 tobacco study was composed of Utah [33.4%], Nevada [23.4%], Montana [19.9%], Colorado [16.4%], and Connecticut [6.9%]).

Minimizing Pre-Intervention Root Mean Squared Prediction Error (RMSPE)

Let X₁ denote a k × 1 vector of pre-intervention characteristics for the treated unit (containing baseline demographic covariates and linear combinations of pre-intervention outcomes Y_1,t for t ≤ T₀), and let X₀ denote the k × J matrix of the same variables for the J donor units. Given a positive semi-definite diagonal predictor-importance matrix V, the optimal donor weights W^*(V) minimize the weighted Euclidean distance:

|| X₁ − X₀ W ||_V = [ (X₁ − X₀ W)^T V (X₁ − X₀ W) ]^1/2

In the outer optimization loop, V is chosen to minimize the pre-intervention Root Mean Squared Prediction Error (RMSPE) of the outcome trajectory across the T₀ pre-treatment periods:

RMSPE_pre = [ (1 / T₀) ∑_t=1^T₀ ( Y_1t − ∑_j=2^J+1 w_j^* Y_jt )² ]^1/2

Once the optimal weights w_j^* are locked using strictly pre-intervention data (t ≤ T₀), the counterfactual outcome at each post-intervention period (t > T₀) and the estimated causal effect τ̂_1t are:

Ŷ_1t(0) = ∑_j=2^J+1 w_j^* Y_jt, and τ̂_1t = Y_1t − Ŷ_1t(0)

In-Space Placebo Tests and the Post/Pre RMSPE Ratio

Because a single treated jurisdiction (N₁ = 1) does not support classical large-sample asymptotic standard errors, Abadie, Diamond, and Hainmueller (2010) introduced exact permutation inference (in-space placebo testing), analogous to Ronald Fisher's randomization test:

1. Reassign the treatment status sequentially to every untreated unit j in the donor pool (j = 2, ..., J + 1), placing the truly treated unit into the donor pool, and compute a placebo synthetic control and placebo gap trajectory τ̂_jt for each donor unit.
2. Because some extreme donor units cannot be well approximated by the convex hull of remaining donors (exhibiting large pre-intervention RMSPE_pre,j), compute for every unit j ∈ {1, ..., J + 1} the ratio of post-intervention RMSPE to pre-intervention RMSPE:

R_j = RMSPE_post,j / RMSPE_pre,j

1. Calculate the exact empirical two-sided permutation p-value as the proportion of all J + 1 units whose ratio R_j is at least as large as that of the treated unit R₁:

p_perm = ( ∑_j=1^J+1 1(R_j ≥ R₁) ) / (J + 1)

Thus, with J = 39 untreated donor states (J + 1 = 40 total units), if the treated state has the single largest post/pre RMSPE ratio (R₁ ranks 1st out of 40), the exact permutation p-value is 1 / 40 = 0.025. Investigators also run in-time placebo tests by artificially backdating the policy date T₀ into the middle of the pre-intervention era to verify that no spurious divergence occurs before actual policy implementation, and leave-one-out sensitivity analyses dropping each positive-weight donor state one at a time.

When the treated unit has the highest or lowest outcome rate in the entire country, it lies outside the convex hull of the donor pool, so no non-negative weights summing to 1 can achieve RMSPE_pre ≈ 0. Recent biostatistical and econometric extensions resolve this boundary limitation:

• Augmented Synthetic Control Method (Ben-Michael, Feller, & Rothstein, 2021): Combines standard simplex SCM weights with a ridge-regularized outcome model to correct small pre-intervention fit discrepancies while bounding extrapolation Penalty.
• Generalized Synthetic Control (Xu, 2017) and Synthetic Difference-in-Differences (Arkhangelsky et al., 2021): Combine unit-specific additive intercepts (allowing a constant baseline level offset, as in DiD) with convex synthetic unit weights and time weights, accommodating both multiple treated units and staggered policy adoption.

Which units must be excluded from the synthetic control donor pool?

Two categories of jurisdictions must be excluded from the donor pool before optimization: (1) any jurisdiction that enacted a similar policy or competing health intervention during the study window (t ∈ [1, T]), and (2) geographically adjacent jurisdictions that experienced spillover (interference) from the treated unit (for example, neighboring states where residents crossed the border to purchase untaxed cigarettes or sugar-sweetened beverages).

Why is overfitting a risk when T₀ is short?

If the pre-intervention window T₀ contains very few time points (for example, T₀ = 3 years) while the donor pool J is large (J = 50 states), the convex optimization can match transient random measurement noise ε_1t rather than true structural factor loadings μ₁. Reliable synthetic control estimation requires a substantial pre-intervention trajectory (typically T₀ ≥ 10 to 20 time periods) and out-of-sample pre-intervention validation splits.

1. Abadie, A., & Gardeazabal, J. (2003). The economic costs of conflict: A case study of the Basque Country. American Economic Review, 93(1), 113–132. https://doi.org/10.1257/000282803321455188
2. Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic control methods for comparative case studies: Estimating the effect of California's tobacco control program. Journal of the American Statistical Association, 105(490), 493–505. https://doi.org/10.1198/jasa.2009.ap08746
3. Bouttell, J., Craig, P., Lewsey, J., Robinson, M., & Popham, F. (2018). Synthetic control methodology as a tool for evaluating population-level health interventions. Journal of Epidemiology and Community Health, 72(8), 673–678. https://doi.org/10.1136/jech-2017-210106
4. Ben-Michael, E., Feller, A., & Rothstein, J. (2021). The augmented synthetic control method. Journal of the American Statistical Association, 116(536), 1789–1803. https://doi.org/10.1080/01621459.2021.1929245