Robust Inference in Models Identified via Heteroskedasticity

Abstract:

Identification via heteroskedasticity, following the approach of Rigobon (2003), is widely used in empirical research. I demonstrate, both theoretically and empirically, that the identification scheme is subject to concerns of weak identification, causing standard inference to be unreliable. This paper offers the methods necessary to confront these issues. I prove conditional identification results and use them to establish conditions under which robust (non-conservative) inference methods for a subset of the parameter vector are valid. I show that these methods are consistently well-sized, unlike the previously best alternative of a projection test, which is dramatically under-sized. I also propose two existing tests for weak identification. I extend the results to impulse response functions based on weakly identified structural estimates, and derive a computationally convenient implementation for confidence intervals. I apply these new methods to Nakamura & Steinsson (2018).

Last updated on 10/22/2018