Robust Inference when Nuisance Parameters may be Partially Identified with Applications to Synthetic Controls

TL;DR

This paper introduces a new inference method for synthetic control estimators that remains valid even when nuisance parameters are partially identified, high-dimensional, and constrained, by orthogonalizing moment conditions.

Contribution

It develops a novel approach combining regularization and orthogonalization to achieve asymptotic normality under complex nuisance parameter conditions.

Findings

Method achieves asymptotic normality despite partial identification.

Applicable to high-dimensional, constrained nuisance parameters.

Verified in synthetic control application.

Abstract

When conducting inference for the average treatment effect on the treated with a Synthetic Control Estimator, the vector of control weights is a nuisance parameter which is often constrained, high-dimensional, and may be only partially identified even when the average treatment effect on the treated is point-identified. All three of these features of a nuisance parameter can lead to failure of asymptotic normality for the estimate of the parameter of interest when using standard methods. I provide a new method yielding asymptotic normality for an estimate of the parameter of interest, even when all three of these complications are present. This is accomplished by first estimating the nuisance parameter using a regularization penalty to achieve a form of identification, and then estimating the parameter of interest using moment conditions that have been orthogonalized with respect to the nuisance parameter. I present high-level sufficient conditions for the estimator and verify these conditions in an example involving Synthetic Controls.