Abstract
Synthetic Control (SC) methods are widely used for policy evaluation but typically struggle with multiple simultaneous treatments. This paper introduces Synthetic Interventions (SI), a framework that generalizes SC by leveraging a low-rank tensor factor model. By modeling data across Time, Units, and Treatments, SI allows researchers to estimate counterfactuals for complex "what-if" scenarios by borrowing strength across different interventions. We provide formal proofs for the consistency of the SI-PCR estimator and validate the approach using the canonical California tobacco study.
Introduction: Beyond Single-Treatment Evaluation
In policy evaluation, the Synthetic Control (SC) method is a foundational tool. For instance, to estimate the impact of a tobacco tax in California, SC constructs a "synthetic California" from a weighted average of states that did not implement the tax. However, standard SC is inherently limited to a single treatment type. It cannot easily handle "data asymmetry"—situations where we wish to learn weights from units under a control setting and apply them to units under a different treatment setting. The central challenge addressed here is how to extend the SC framework to multiple treatments while maintaining rigorous theoretical guarantees.
Methodology: The Tensor Perspective
The core innovation of Synthetic Interventions is the transition from a matrix factor model to a tensor factor model. Instead of a 2D grid (Time $\times$ Units), the data is modeled as a 3D tensor: Time $\times$ Units $\times$ Treatments.
Latent Factor Invariance
The model assumes that each unit (e.g., a state) possesses a set of latent "unit factors" that remain invariant across time and different treatments. This invariance allows the model to map relationships learned under one treatment (e.g., the status quo) to another (e.g., a tax hike).
The SI-PCR Estimator
The authors propose the SI-PCR (Principal Component Regression) estimator, which operates in two stages:
- Model Identification: Pre-intervention data is used to find optimal weights that represent a target unit as a combination of "donor" units.
- Point Estimation: These weights are applied to the observed outcomes of donor units under the new treatment to predict the target unit's counterfactual.
Crucially, the authors provide formal proofs for the consistency and asymptotic normality of this estimator, enabling the construction of valid confidence intervals for multi-treatment contexts.
Key Insights: Causal Inference as Tensor Completion
By framing causal inference as a "missing data" problem in a 3D tensor, the goal becomes tensor completion. If every entry in the Time $\times$ Unit $\times$ Treatment tensor were known, causal inference would be trivial. Since only a sparse subset is observed, SI uses low-rank structures to fill the gaps.
Furthermore, the paper links causal inference to covariate shift. For the model to generalize from control to treatment, the latent factors of the target unit must lie within the "span" of the donor units' factors under the new treatment. The SI-PCR estimator effectively de-noises the data by focusing on these stable, low-rank latent structures.
Empirical Validation: The Tobacco Study
The framework was evaluated using synthetic simulations and a revisit of the 1988 California Proposition 99 case study. This dataset tracks cigarette sales across 50 U.S. states over 31 years, categorized into three treatment slices: status quo, anti-tobacco programs, and tax increases.
Key Results
- Consistency: Theorem 2 proves that estimation error vanishes as data size increases, provided the pre-intervention period is sufficiently long ($T_0 = \omega(N_d^{1/3})$).
- Policy Impact: In the tobacco study, the SI-PCR estimator found that anti-tobacco programs and tax hikes had remarkably similar effects, both significantly reducing consumption compared to the status quo.
- Accuracy: A leave-one-out analysis showed a normalized prediction error of approximately 0.07 for the "raised taxes" treatment, outperforming the baseline "status quo" error (0.105).
Limitations and Future Work
While promising, the current theoretical results assume independent sub-Gaussian noise, whereas real-world panel data often features correlated shocks. Additionally, the method relies on the "span" assumption; if a target unit is a unique outlier, the synthetic intervention may not converge. Future research could explore "design-based" properties of SI in randomized experiments versus observational data.
Summary updated