Year: 2024
Type: preprint
Abstract: The Synthetic Control (SC) method is widely employed in estimating causal treatment effects in observational studies.Typically, it synthesizes the counterfactual of the treated unit by employing a data-driven weighted average of the remaining units in the post-treatment period.These weights aim to minimize the distance between the treated unit and its counterfactual in the pre-treatment period.To prevent overfitting of the pre-treatment data, the original approach introduced by Abadie, Diamond, and Hainmueller (ADH) imposes a constraint wherein all weights must be weakly positive and sum up to one.Expanding upon Doudchenko and Imbens [2016], we introduce REGSC, an alternative regularized synthetic control approach.This approach involves shrinking individual coefficients toward zero and the sum of coefficients toward one.It amalgamates the benefits of the original SC method and the elastic net, offering a closed-form solution with tunable hyperparameters.The estimator also allows for a straightforward Bayesian representation, providing advantages, particularly in quantifying estimation uncertainty through Bayesian credibility intervals.By appending the donor pool with lagged values of the donors and the counterfactual, the REGSC framework is extended to dynamic contexts.Our dynamic version of the REGSC estimator accommodates nonstationary and cointegrated time series, a case where the original SC method is inconsistent.To assess the relative performance of our estimation approach in static and dynamic data-generating processes we conduct a variety of Monte Carlo experiments experiments and apply it to existing empirical datasets.Our results suggest that the (dynamic) REGSC estimator outperforms other SC methods previously proposed in the literature.
Authors: Breitung, Jörg, Bolwin, Lennart, Töns, Justus
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6998498778