Macroeconomic data are largely non-experimental. Thus, causal inference in macroeconomics is typically based on assumptions about what aspects of the variation in the data are exogenous. This presents two major challenges, which this research addressed directly. First, few such assumptions are generally accepted. Hence, conclusions about the causal effects of macroeconomic policy, and, therefore, the resulting policy prescriptions, can vary significantly depending on the underlying identifying assumptions. Second, conditional on any set of assumptions, identification of causal effects is often weak because there is little relevant variation in the data. Therefore, inference methods must accurately reflect the amount uncertainty in the data about the effects of policy, for otherwise, researchers working on the same problem can obtain seemingly contradictory results, leading to empirical puzzles. In the words of the nineteenth century philosopher Carveth Read, “it is better to be vaguely right than precisely wrong.”
This project proposed to tackle the above two challenges via two parallel lines of enquiry that explored new sources of identification and developed the requisite econometric methods that characterize the information in the data accurately and efficiently.
The objective of the first line of enquiry has been to explore the potential to identify the causal effects of macroeconomic policy by comparing the behaviour of the economy across different policy regimes. The idea is that regime changes may act as ‘quasi-experiments’ that generate variation can help identify causal effects of policy over time. Policy regime changes may arise for various reasons, such as shifts in policy objectives, or occasionally binding constraints on the policy instruments. The latter possibility had not been considered in the literature before. A leading example of an occasionally binding is the so-called ‘zero lower bound’ (ZLB) constraint on the nominal interest rate that is used by a central bank as the primary policy instrument for the conduct of monetary policy. The ZLB constraint on interest rates first became binding in Europe and the United States following the Great Financial crisis of 2008, and again very recently during the coronavirus pandemic, though it had occurred in Japan much earlier following the Asian crisis of the late 1990s. The ZLB has been a major challenge for policy makers because it has constrained their ability to conduct monetary policy via the ‘conventional’ and relatively well-understood monetary policy channel, which is the variation of the nominal interest rate, and has forced them to implement previously unexplored so-called ‘unconventional’ monetary policies, such as quantitative easing, whose effectiveness had not been studied before. In this project, I proposed to turn this problem on its head and use it as an opportunity to learn about the causal effects of conventional and unconventional policy.
The objective of the second line of enquiry has been to contribute to the ongoing research on developing methods of inference that are robust to the problem of ‘weak identification’, that is, methods that deliver accurate and reliable inference even when there is little relevant information in the data to identify the causal effects of interest. Weak identification is pervasive in economics and threatens the validity of inference under any identification scheme. This problem is therefore of central importance for causal inference using non-experimental data in economics and other fields. As a result, a burgeoning literature has developed to tackle this problem over the last two decades and has produced some major breakthroughs. However, a limitation of existing methods has been that in many cases they may sacrifice power for robustness, i.e. they limit the risk of erroneous conclusions drawn from statistical tests at the cost of giving answers that are potentially too vague. The objective of this line of enquiry has been to explore whether the statistical power of tests in specific but widely applicable models can be improved while maintaining the reliability of inference when there is very little information in the data.
