Synthetic Regressing Control

Technical Summary: Synthetic Regressing Control (SRC)

Problem Statement
The Synthetic Control (SC) method is a standard approach for estimating treatment effects in panel data with a single treated unit and a modest number of control units. It constructs a "synthetic control" by finding a weighted average of control units that closely matches the treated unit's pre-treatment outcomes. However, the paper identifies two critical oversights in the standard SC methodology that lead to suboptimal weights and poor extrapolation:

1. Interpolation Error: The standard SC method attempts to match pre-treatment outcomes directly. If the linear combination of control units (constrained to non-negative weights summing to one) cannot closely approximate the treated unit's pre-treatment trajectory, significant interpolation error occurs. This is particularly problematic under factor models where the treated unit's factor loadings lie outside the convex hull of the control units' loadings.
2. Noise and Constraint Oversights: The standard constraint that weights must sum to one ($\sum w_j = 1$) is designed to minimize bias in the mean but overlooks the impact of noise. The paper demonstrates that this constraint fails to minimize the variance component of the prediction error. Specifically, forcing the sum of weights to equal 1 does not optimize the trade-off between bias and variance in the presence of idiosyncratic errors, leading to suboptimal mean-squared prediction error.

Methodology: Synthetic Regressing Control (SRC)
The proposed Synthetic Regressing Control (SRC) method addresses these issues through a two-stage procedure that decouples the alignment of units from the synthesis of weights.

1. Unit Regression (Pre-processing):
   Instead of directly matching raw outcomes, SRC first performs a univariate linear regression for each control unit $j$ against the treated unit using pre-treatment data.

   • For each control unit $j$, a coefficient $\hat{\theta}_j$ is estimated to minimize the squared distance between the treated unit's pre-treatment outcomes and a scaled version of the control unit's outcomes.
   • This generates "regressed controls" $\tilde{Y}_{jt}(0) = \hat{\theta}_j(Y_{jt} - \bar{y}_j)$, which effectively align the pre-treatment trajectories of the control units with the treated unit, reducing interpolation error.

2. Synthesis with Unbiased Risk Estimation:
   The SRC estimator is constructed as a weighted average of these regressed controls. The weights $w$ are determined by minimizing an unbiased risk estimator (specifically, a Mallows' $C_p$-style criterion) rather than satisfying the strict simplex constraint ($\sum w_j = 1$).

   • Objective Function: The method minimizes $\| \hat{y}_1(w) - y_1 \|^2 + 2\hat{\sigma}^2 \sum w_j$, where $\hat{\sigma}^2$ is an estimate of the error variance.
   • Constraint: The weights are constrained to be non-negative ($w_j \ge 0$) but are allowed to sum to a constant $c$ (where $c$ is data-driven and related to the noise level), rather than strictly summing to 1. This relaxation allows the method to account for model noise and prevents the "over-fitting" of the noise component that occurs when forcing the sum to 1.
   • High-Dimensional Extension: When the number of control units $J$ exceeds or approaches the number of time periods $T_0$, the paper proposes a screening step using Sure Independent Ranking and Screening (SIRS) to select a subset of active units before applying the SRC algorithm.

Key Contributions

• Theoretical Framework: The paper introduces a working model where the treated unit's potential outcome is a linear combination of regressed control units plus an error term. This framework explicitly separates the alignment of units (via regression coefficients) from the synthesis of units (via weights).
• Asymptotic Optimality: The authors prove that the SRC estimator, using weights derived from the unbiased risk estimator criterion, is asymptotically optimal. Specifically, the loss of the SRC estimator converges to the minimum loss achievable by the "infeasible best possible" synthetic estimator (the oracle) as the number of pre-treatment periods increases.
• Addressing SC Limitations: The method theoretically demonstrates that relaxing the $\sum w_j = 1$ constraint in favor of a penalized risk minimization reduces the impact of noise on the prediction error, a limitation inherent in standard SC.
• Handling Non-Linearity and Heterogeneity: Through examples involving linear factor models, nonlinear factor models, and autoregressive models, the paper shows that unit regression can identify and down-weight irrelevant control units (e.g., setting weights to zero in nonlinear factor models where controls are unrelated to the treated unit), thereby improving robustness.

Results

• Simulation Studies: Extensive Monte Carlo simulations based on factor models (including settings with fixed time effects, heterogeneous factor loadings, and heteroscedastic noise) show that SRC consistently achieves lower mean-squared prediction error compared to existing methods, including the original SC, demeaned SC, augmented SC, generalized SC, and constrained Lasso.
• Empirical Application: The method is applied to the economic costs of conflicts in the Basque region of Spain. The results indicate that SRC provides a better fit for the pre-treatment period and yields more precise counterfactual estimates compared to alternative estimators.

Significance and Claims
The paper claims that SRC offers a simple yet effective improvement over the standard Synthetic Control method by addressing the fundamental issues of interpolation error and noise sensitivity.

• Modest Claims on Inference: The authors explicitly state that while the SRC estimator achieves asymptotic optimality in terms of prediction loss, it does not currently provide an unbiased estimator for the treatment effect itself. They note that deriving a debiased estimator for valid statistical inference (e.g., confidence intervals) using projection theory is a promising direction for future work but lies beyond the scope of the current paper.
• Interpretability: The method retains the interpretability of the SC approach (non-negative weights) while allowing the regression coefficients to handle the scaling and alignment, effectively managing extrapolation without relying on the restrictive unit-sum constraint.
• Practical Utility: The inclusion of a screening mechanism (SIRS) makes the method applicable to high-dimensional settings where the number of control units is large relative to the time series length, a common challenge in modern panel data applications.