Technical Summary: Targeted Synthetic Control Method

Problem Statement

The Synthetic Control Method (SCM) is a standard approach for estimating causal effects in panel data settings involving a single treated unit and multiple untreated control units. The core objective is to construct a counterfactual outcome for the treated unit as a convex combination (weighted average) of control units, calibrated to match the treated unit's pre-treatment trajectory.

However, classical SCM faces two primary limitations:

1. Sensitivity to Finite-Sample Error: The estimator relies entirely on pre-treatment matching. If the pre-treatment fit is imperfect due to finite sample sizes, the resulting weights can be poorly estimated, leading to biased counterfactual predictions in the post-treatment period.
2. Lack of Flexibility: Classical SCM typically uses linear matching without leveraging flexible machine learning (ML) models to capture complex outcome regressions.

Existing extensions, such as the Augmented SCM, attempt to incorporate flexible ML outcome regressions to reduce bias. However, they do so by adding an additive residual correction term to a plug-in prediction. This approach breaks the convex combination structure, meaning the final counterfactual estimate is no longer a weighted average of observed control outcomes. Consequently, the estimates can become unbounded (falling outside the range of observed control outcomes), and the weights lose their direct interpretability as relative contributions of control units.

Methodology: Targeted Synthetic Control (TSC)

The authors propose the Targeted Synthetic Control (TSC) method, a two-stage estimator inspired by Targeted Maximum Likelihood Estimation (TMLE). TSC aims to directly estimate the counterfactual outcome while preserving the interpretability and boundedness of the synthetic control structure.

Two-Stage Procedure

1. Nuisance Component Estimation:

   • Initial Weights: TSC begins by obtaining an initial set of synthetic-control weights ($\hat{w}^0$) using a standard matching procedure (e.g., minimizing the distance between the treated unit's pre-treatment history and a weighted combination of control units).
   • Outcome Regression: A flexible machine learning model is trained on the control units to estimate the outcome regression function $\hat{m}_{\tilde{t}}(x) = E[Y_{\tilde{t}} | X=x, A=0]$. This model predicts outcomes based on covariates (including pre-treatment history).

2. Targeted Update (Debiasing):

   • Instead of adding a correction term to the prediction, TSC performs a one-dimensional targeted update on the initial weights.
   • Weight-Tilting Submodel: The initial weights are updated via an exponential tilting (softmax) submodel:
     $$\hat{w}_j(\varepsilon) = \frac{\hat{w}^0_j \exp(\varepsilon S_j)}{\sum_{k=2}^N \hat{w}^0_k \exp(\varepsilon S_k)}$$
     where $S_j$ represents centered residuals (e.g., $S_j = \hat{m}_{\tilde{t}}(X_j) - \sum \hat{w}^0_k \hat{m}_{\tilde{t}}(X_k)$). This ensures the updated weights remain non-negative and sum to one (preserving the simplex constraint).
   • Calibration: The parameter $\hat{\varepsilon}$ is chosen such that the weighted residuals of the control units sum to zero:
     $$\sum_{j=2}^N \hat{w}_j(\hat{\varepsilon}) \left( Y_{j\tilde{t}} - \hat{m}_{\tilde{t}}(X_j) \right) = 0$$
   • Final Estimator: The final counterfactual is computed as the convex combination of observed control outcomes using the updated weights $\hat{w}^* = \hat{w}(\hat{\varepsilon})$:
     $$\hat{\psi}^{TSC}_{\tilde{t}} = \sum_{j=2}^N \hat{w}^*_j Y_{j\tilde{t}}$$

Key Contributions

The paper claims three primary contributions:

1. Novel Link to TMLE: It is the first work to link SCM to a TMLE-style targeted updating perspective while retaining interpretable synthetic-control weights. Unlike standard TMLE which updates the outcome regression, TSC updates the weights.
2. Bounded and Interpretable Estimation: The method ensures the final counterfactual is strictly a convex combination of observed control outcomes. This guarantees that estimates remain within the convex hull of the control data (boundedness) and that weights retain their interpretation as relative contributions.
3. Robustness and Accuracy: By using flexible ML models and enforcing the residual-correcting term to zero through the targeted update, TSC mitigates bias arising from imperfect pre-treatment weight estimation and model misspecification.

Experimental Results

The authors evaluated TSC against Classical SCM, Plug-in estimators, and Augmented SCM using both synthetic and real-world data.

• Synthetic Data: Experiments covered continuous and binary outcomes with various data-generating functions (linear, hinge, quadratic, time-varying).

  • Accuracy: TSC consistently achieved the lowest Root Mean Squared Error (RMSE) across all settings and prediction horizons. For binary outcomes, TSC reduced RMSE by approximately 18% compared to Augmented SCM.
  • Boundedness: In binary outcome experiments, Augmented SCM produced counterfactual estimates outside the $[0, 1]$ range in significant percentages (up to 52% violation in some scenarios). In contrast, TSC achieved a 0.00% violation rate, strictly adhering to the bounds of the data.
  • Interpretability: Visual analysis confirmed that TSC weights remained non-negative and sum to one, whereas Augmented SCM weights lost this property.

• Real-World Data:

  • New Hampshire Turnout (1996): TSC produced a stable counterfactual trajectory that closely matched pre-treatment history, suggesting a large positive treatment effect. Classical SCM showed downward deviation, and Augmented SCM showed upward deviation.
  • California Tobacco Control (1988): TSC generated a counterfactual within the range of control outcomes, indicating a substantial negative treatment effect (reduced cigarette sales) consistent with historical findings. Augmented SCM and Plug-in estimators produced less plausible post-treatment trajectories.

Significance and Claims

The paper positions TSC as a method that resolves the trade-off between flexibility and interpretability in synthetic control methods.

• Theoretical Significance: It provides a rigorous framework for debiasing synthetic control weights without sacrificing the convex combination structure, a property often lost in augmented approaches.
• Practical Significance: By ensuring boundedness, TSC prevents the "unbounded extrapolation" problem inherent in additive correction methods, which is particularly critical for bounded outcomes (e.g., probabilities, binary events).
• Future Potential: The authors suggest the methodology could be applicable to other settings, such as single-arm trials in medicine utilizing external or historical controls, though they do not elaborate on specific implementations for these cases.

The authors maintain a modest tone, noting that while TSC improves upon state-of-the-art baselines in their experiments, the method relies on the existence of synthetic-control weights (Assumption 3.1) and standard causal inference assumptions (SUTVA, no anticipation, shared untreated outcome model).