B1832
Title: On the assumptions and misspecifications of synthetic controls
Authors: Claudia Shi - Columbia University (United States) [presenting]
Abstract: Synthetic control (SC) methods are widely used to estimate causal effects from observational data, by approximating a treated unit's counterfactual outcomes as a weighted combination of control units. Valid causal inference hinges on the critical linearity assumption - that the treated unit can be written as a linear combination of controls. The identifiability and robustness of SC are examined. First, the problem is reformulated using more granular individual-level data, revealing how linearity emerges naturally from this flexible model. This highlights new strategies for sample selection to improve identifiability. Building on this fine-grained model, the misspecification error is theoretically bound when linearity is violated. The bounds show small misspecifications induce small errors. Leveraging this insight, new SC estimators are developed that minimize misspecification by incorporating additional demographic data. The validity and usefulness of these estimators are demonstrated on synthetic and real-world data.