Credible Inference for Empirical Macroeconomics

Introduction

Following the credibility revolution, macroeconomists have sought plausibly exogenous instruments and other sources of variation to identify causal effects. Given the complex nature of the macroeconomy, characterised by simultaneous causality and intertemporal dependence, this is a high bar.

Challenges in Identification

Thus, in the pursuit of exogenous variation, researchers often use minor sources of variation or subtle features of the data to identify the effects of interest. When the variation exploited is modest, “weak identification” can arise. In practice, this means that estimators are no longer asymptotically normal, so standard techniques for statistical inference – conducting hypothesis tests or constructing confidence intervals – are invalid.

Current Research Landscape

While this likely occurs in much empirical research in macroeconomics, few papers acknowledge these issues, partially because there are rarely appealing options to address them.

Proposal Overview

This proposal provides attractive options for researchers to combat weak identification in macroeconometric models.

Avoiding Weak Identification

1. Novel Frameworks: It offers the possibility to avoid weak identification in the first place, via novel frameworks to exploit instrumental variables in panel and time series data.
2. Rich Information Extraction: These frameworks extract richer information from a given instrument and expand the set of admissible instruments.

Tools for Confidence Sets

Next, I provide tools to construct confidence sets for dynamic causal effects, a key object of interest, that are valid regardless of how strong the identifying variation is.

1. Existing Approaches: Existing approaches produce confidence sets that are conservative – too large.
2. Improved Models: I first consider models identified using instrumental variables, improving both computational burden and performance relative to frontier methods.
3. General Sources of Variation: Finally, I consider models identified using more general sources of variation, and, working identification scheme by scheme, provide performance gains over leading methods for confidence sets.

Conclusion

I thus facilitate credible inference to match credible identification strategies.