The synthetic control method (SCM) estimates causal effects in panel data with a single-treated unit by constructing a counterfactual outcome as a weighted combination of untreated control units that matches the pre-treatment trajectory. In this paper, we introduce the targeted synthetic control (TSC) method, a new two-stage estimator that directly estimates the counterfactual outcome. Specifically, our TSC method (1) yields a targeted debiasing estimator, in the sense that the targeted updating refines the initial weights to produce more stable weights; and (2) ensures that the final counterfactual estimation is a convex combination of observed control outcomes to enable direct interpretation of the synthetic control weights. TSC is flexible and can be instantiated with arbitrary machine learning models. Methodologically, TSC starts from an initial set of synthetic-control weights via a one-dimensional targeted update through the weight-tilting submodel, which calibrates the weights to reduce bias of weights estimation arising from pre-treatment fit. Furthermore, TSC avoids key shortcomings of existing methods (e.g., the augmented SCM), which can produce unbounded counterfactual estimates. Across extensive synthetic and real-world experiments, TSC consistently improves estimation accuracy over state-of-the-art SCM baselines.
Figures & tables
Figure 1: Synthetic control setting . Panel data with outcomes for single-treated unit ( red ) and multiple control units ( green ). Treatment is assigned at T0 , so that the pre-treatment history of the treated unit is used to learn synthetic-control weights based on other pre-treatment histories of control units ( gray ), while the post-treatment period is used to estimate the treatment effect between the treated unit and the synthetic control (dark arrow).
Aspects
Classical SCM [ 2 , 4 ]
Augmented SCM [ 8 ]
TSC (ours)
Core idea
Pre-treatment balancing weights
Outcome-model augmentation
TMLE-style weights tilting
Counterfactual form
Convex combination of control outcomes
Prediction + additive adjustment
Convex combination of control outcomes
ML model
✗ ( weights only )
✓ ( regression )
✓ ( regression )
Interpretable weights
✓ ( weighted average )
✗ ( mixture format )
✓ ( weighted average )
Boundedness (e.g., binary)
✓ ( convex hull )
✗ ( prediction and additive term )
✓ ( convex hull )
Table 1: Overview of key synthetic control estimators. The table compares classical SCM, augmented SCM, and TSC in terms of how they construct the counterfactual and what properties are preserved by the estimator’s final form.
Figure 2: Overview of TSC . Our method consists of two stages: 1 a nuisance component estimation and 2 a targeted update for debiasing. In the nuisance stage, we (i) compute initial synthetic-control weights w^ by matching pre-treatment histories and (ii) fit an outcome-regression m^t~(x) on control units using any machine-learning backbone. In the debiasing stage, we construct residual scores Sj of control units and perform a one-dimensional exponential-tilting update w~(ε) . The parameter ε^ is obtained by solving the condition f(ε^)=0 . The resulting targeted weights w^⋆=w~(ε^) yield the final counterfactual ψ^t~=∑j=2Nw^j⋆Yjt~ (and thus the treatment effect) while preserving interpretable synthetic-control weights.
Prediction horizon
Methods
Linear
Hinge
Quadratic
Time-varying
1
Classical SCM
0.922±0.146
0.988±0.163
1.030±0.186
0.888±0.126
Plug-in estimator
1.053±0.245
1.118±0.292
1.434±0.401
0.987±0.212
Augmented SCM
1.111±0.245
1.186±0.292
1.458±0.401
1.049±0.212
TSC ( ours )
0.892±0.134
0.924±0.139
1.022±0.188
0.879±0.126
5
Classical SCM
0.975±0.233
0.964±0.230
1.044±0.226
0.935±0.236
Plug-in estimator
0.993±0.346
1.001±0.364
1.152±0.422
1.012±0.365
Table 2: Results for experiments with synthetic data. Mean ± standard error across five runs, reported for prediction horizons of 1, 5, 10 steps ahead. ⇒ Our TSC performs best across all settings.
Figure 3: Insights of weights. Bars show the weights of control units used to construct the counterfactual ( top row : continuous; bottom row : binary). Shown are the initial synthetic control weights (light colors) and the final weights (dark colors).
Bound
Method
Linear
Hinge
Quadratic
Time-varying
Upper bound
Augmented SCM
3.33%
13.33%
20.00%
26.67%
TSC (ours)
0.00%
0.00%
0.00%
0.00%
Lower bound
Augmented SCM
13.33%
40.00%
13.33%
13.33%
TSC (ours)
0.00%
0.00%
0.00%
0.00%
Table 3: Percentage of counterfactual estimations outside [0,1] on binary outcomes. ⇒ TSC achieves 0.00% violation rate; augmented SCM does not.
Figure 4: Real-world experiments. The trajectories of the control units are shown in gray , and the treated unit in red . We plot counterfactual estimates from augmented SCM ( blue ), classical SCM ( orange ), the plug-in estimator ( purple ), and TSC ( green ). The vertical dashed line indicates the treatment time T0 . TSC matches the pre-treatment history closely in both datasets.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Table 8
Method
h=1
h=5
h=10
Augmented SCM
1.24
1.18
0.61
TSC ( ours )
0.53
0.54
0.49
Appendix
Table 6: Oracle gain in RMSE under the latent-factor DGP across different prediction horizons.
Method
Lag 1
Lag 5
Lag 10
GSC
10.612±0.405
10.680±0.468
10.670±0.478
MC
10.668±0.401
10.667±0.418
10.638±0.413
SyncTwin
10.921±1.341
9.825±1.342
9.925±1.343
TSC ( ours )
10.150±1.286
9.733±1.127
9.923±1.328
Appendix
Table 7: Comparison with additional baselines under the synthetic quadratic DGP.
Method
Lag 1
Lag 5
Lag 10
TSC + GSC mt(X)
9.991±1.656
10.015±1.662
10.007±1.659
TSC + MC mt(X)
9.267±1.901
9.270±1.902
9.259±1.898
TSC + SyncTwin mt(X)
9.312±0.609
9.599±0.614
10.184±0.596
Appendix
Table 8: Backbone ablation of TSC under the synthetic quadratic DGP.
Data Science Institute Columbia University New York, NY 10027, USA · Department of Computer Science and Department of Statistics Columbia University New York, NY 10027, USA