Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
Synthetic Control with Disaggregated Data
Authors: Lea Bottmer,
Presenting Author: Lea Bottmer*
The synthetic control estimator is widely used to evaluate aggregate-level policies, but researchers increasingly face settings with rich, disaggregated data (e.g., county-level outcomes within states) that raise new questions about aggregation choice. Existing approaches incorporate such data by estimating separate synthetic controls for each disaggregated treated unit, enlarging the donor pool with disaggregated control units, or both. These strategies can improve fit but also amplify noise, with little guidance on how to balance these trade-offs. This paper develops a general framework for synthetic control with disaggregated data that nests the classical synthetic control estimator and other existing approaches. Within this framework, I propose a multi-level SC (mlSC) estimator that formalizes the aggregation choice as a data-driven regularization problem. The estimator flexibly regularizes toward the classical synthetic control estimator while exploiting additional variation from the disaggregated data. In simulations calibrated to four empirical settings, mlSC matches or outperforms existing approaches. Two applications—Minnesota’s cigarette tax and minimum wage effects on teen employment—illustrate its practical value.