I’ve spent the last month or something reading about causal identification, more particularly, on the “new experimental” school and (with more interest) their critics. A nice reading list can be found here and I liked these two issues of the JEP and the JEL .
If you keep reading this blog, you will probably discover that I tend to lie toward the structuralist side of the debate and that I have little sympathy towards the “empiricist” view of causality. I tended to think that the main motivation for this had to do with the (a) The academic environment where I was raised (b) The role that formal theory (or even informal game theoretic settings) play in my thought process.
It was not until recently when I thought of a third factor: at heart, I’ve always considered myself a macroeconomist, and modern macro is basically structural at its very root. I think I started to realized this when I read Chris Sims’ piece in the JEP (I saw Chris Sims at a conference last year and watching his mind in action is an amazing experience). I thought, however, of writing this piece written by John Cochrane on economic writing (the reading list is here) The paper has very interesting lines, but the paragraph that kept my attention is the following:
Are you sure causality doesn’t run from y to x, or from z to y and x simultaneously? Think of the obvious reverse-causality stories.
Example: You can also think about the last example as causality: Do interest rates cause changes in housing demand or vice versa (or does the overall state of the economy cause both to change)?
This is the typical problem you are used to deal in macroeconomics. When you do general equilibrium macro, you typically deal with general equilibrium setting where everything depends on everything, that is, everything is simultaneous and causality is never unidirectional. In the Rubin identification framework, most papers are written as single equations models where, under some assumptions regressions identify causal relationships between observed variables. Most microeconometric models rule out, by assumption, general equilibrium effects (pg 11 assumption A-1).
In structural vector autoregression, restrictions based on substantive economic reasoning are placed on a multivariate time series model that allows interpretation of some functions of its parameters as policy effects. Practitioners of this method impose identifying restrictions parsimoniously, often leaving most parameters of the model without a behavioral interpretation, and usually leaving the fit of the model to the multivariate time series data as good as that of an unrestricted reduced form model.
In the last few years, starting with the work of Smets and Wouters (2003), models with more complete interpretations than the structural vector autore- gressions, which fifit nearly as well as structural vector autoregressions, have been estimated. These models, called dynamic stochastic general equilibrium models, make much stronger assumptions than the structural vector autoregressions, but they reproduce the implications of the structural vector autogressions for the effects of monetary policy. The fact that the models match in this respect increases confifidence that the dynamic stochastic general equilibrium models are not getting their estimates of monetary policy effects mainly from their strong assumptions.
In both sorts of models, you need to rely on many equations because, due to general equilibrium effects, causality is multidirectional. The birth of modern macro is precisely “Lucas’ Critique” which is also the birth of structural econometrics. Lucas’ challenge was that nothing could be inferred about the relationship between variable under a different policy regime because this could be specific to the current regime. You had, instead, to focus on those parameter that would be plausibly policy invariant: the technology available, the taste and preferences of individuals, etc.
Moreover in DSGE’s, the structural, policy invariant parameters estimated that drive every other variable often have a straightforward behavioral interpretation and the mixing of the theoretical mechanism and the behavior of variable occurs smoothly. If you think on how to estimate the effect of a certain educational policy, someone used to deal with OLG model feels it makes sense to think in terms of a life cycle model where educational choices are part of the career of individuals and the anatomy of the causal mechanism -choices, interactions, etc, emerges naturally and explicitly from the model. Instead, an “empiricist” would try to find a pseudo experimental framework and evaluate some statistical relation between the treatment and the outcome and then turn to some story that seems convincing.