Is Completeness Necessary? Estimation in Nonidentified Linear Models

Abstract

We show that estimators based on spectral regularization converge to the best approximation of a structural parameter in a class of nonidentified linear ill-posed inverse models. Importantly, this convergence holds in the uniform and Hilbert space norms. We describe several circumstances when the best approximation coincides with a structural parameter, or at least reasonably approximates it, and discuss how our results can be useful in the partial identification setting. Lastly, we document that identification failures have important implications for the asymptotic distribution of a linear functional of regularized estimators, which can have a weighted chi-squared component. The theory is illustrated for various high-dimensional and nonparametric IV regressions.

Keywords: completeness condition, weak identification, nonparametric IV regression, high-dimensional regressions, spectral regularization

JEL Classification: C14, C26