Stephen Williamsonが引用したEconomicDynamicsのルーカスインタビューが面白い(H/T Economist's View)。

ED: If the economy is currently in an unusual state, do micro-foundations still have a role to play?
RL: "Micro-foundations"? We know we can write down internally consistent equilibrium models where people have risk aversion parameters of 200 or where a 20% decrease in the monetary base results in a 20% decline in all prices and has no other effects. The "foundations" of these models don't guarantee empirical success or policy usefulness.
What is important---and this is straight out of Kydland and Prescott---is that if a model is formulated so that its parameters are economically-interpretable they will have implications for many different data sets. An aggregate theory of consumption and income movements over time should be consistent with cross-section and panel evidence (Friedman and Modigliani). An estimate of risk aversion should fit the wide variety of situations involving uncertainty that we can observe (Mehra and Prescott). Estimates of labor supply should be consistent aggregate employment movements over time as well as cross-section, panel, and lifecycle evidence (Rogerson). This kind of cross-validation (or invalidation!) is only possible with models that have clear underlying economics: micro-foundations, if you like.

This is bread-and-butter stuff in the hard sciences. You try to estimate a given parameter in as many ways as you can, consistent with the same theory. If you can reduce a 3 orders of magnitude discrepancy to 1 order of magnitude you are making progress. Real science is hard work and you take what you can get.

"Unusual state"? Is that what we call it when our favorite models don't deliver what we had hoped? I would call that our usual state.

「ミクロ的基礎付け」? 人々の危険回避度のパラメータが200であるような、あるいはマネタリーベースの20%の減少がすべての物価を20%低下させてそれ以外には何の影響ももたらさないような内的整合性を有する均衡モデルを構築できることは分かっている。そうしたモデルの「基礎付け」は、実証面での成功や政策ツールとしての有用性を保証するわけではない。
「通常でない状態」? それはお気に入りのモデルがお望みの結果を出してくれないことを指すのかね? 私に言わせれば、それは通常の状態だね。


This I completely and utterly agree with. The power of microfoundations is that they allow unification. They give you the hope of explaining many phenomena in terms of one underlying phenomenon. In fact, there really should be no conceptual difference between the terms "microfoundations" and "unification". Microfounded models - when they work -are inherently more useful and powerful than equivalent "ad-hoc" models.

Of course, as Lucas points out, this makes microfounded models inherently more vulnerable to falsification than other models. An "ad-hoc" non-microfounded model only has to explain one thing, but a microfounded model has to explain many things. This means that as soon as the microfounded model fails to explain any one of the things it purports to explain, you have to conclude that the model is flawed, and look for a better model. (In practice of course, not everyone actually does this when they ought to...)