Synthetic Control Methods and Big Data

TL;DR

This paper introduces a flexible framework for synthetic control methods that integrates machine learning techniques to improve counterfactual prediction in macroeconomic policy analysis, especially with large control pools.

Contribution

It develops a general framework unifying various synthetic control and machine learning approaches, optimizing bias-variance tradeoff for better counterfactual estimation.

Findings

Machine learning methods outperform traditional approaches with large control pools.

Revised estimates show effects of economic liberalization on growth not previously detected.

Analysis of bank responses to capital requirements using extensive bank data.

Abstract

Many macroeconomic policy questions may be assessed in a case study framework, where the time series of a treated unit is compared to a counterfactual constructed from a large pool of control units. I provide a general framework for this setting, tailored to predict the counterfactual by minimizing a tradeoff between underfitting (bias) and overfitting (variance). The framework nests recently proposed structural and reduced form machine learning approaches as special cases. Furthermore, difference-in-differences with matching and the original synthetic control are restrictive cases of the framework, in general not minimizing the bias-variance objective. Using simulation studies I find that machine learning methods outperform traditional methods when the number of potential controls is large or the treated unit is substantially different from the controls. Equipped with a toolbox of approaches, I revisit a study on the effect of economic liberalisation on economic growth. I find effects for several countries where no effect was found in the original study. Furthermore, I inspect how a systematically important bank respond to increasing capital requirements by using a large pool of banks to estimate the counterfactual. Finally, I assess the effect of a changing product price on product sales using a novel scanner dataset.